AI Patent Watch

Granted US AI patents across the 33 active filers, by grant date. The final bar is the current quarter, still in progress. Grants lag filings by roughly two years, so this measures issued IP rather than current research.

• Sep 8 UnitedHealth
  Temporal sequence causal transformer machine learning model
  filed 2023
• Sep 8 UnitedHealth
  Systems and methods for predictive analyses with machine learning systems
  filed 2024
• Sep 8 UnitedHealth
  Machine learning techniques for predicting and ranking typeahead query suggestion keywords based on user click feedback
  filed 2023
• Sep 8 Allstate
  Adaptable On-Deployment Learning Platform for Driver Analysis Output Generation
  filed 2025
• Sep 8 USAA
  Virtual coaching assistant
  filed 2021
• Sep 8 CVS Health
  Machine learning based data management
  filed 2024
• Sep 1 UnitedHealth

  Machine-learning based irrelevant sentence classifier

  A text-comprehension system that uses a machine-learning classifier to flag and strip irrelevant sentences from a narrative, producing a pertinent summary and output indicium that condenses clinical or claims narratives before downstream processing.

  filed 2022 ▾

  The system receives a narrative made of one or more sentences and applies a machine-learning irrelevant-sentence classifier, trained on automatically generated data tailored to a specific circumstance, to score each sentence's relevance. Sentences judged irrelevant are removed to form a pertinent summary, from which the system generates an output indicia data object describing the narrative. By discarding noise before analysis, the approach supports faster and more accurate reading of long free-text records such as medical notes or claim descriptions. The filer is effectively protecting an automated way to pre-clean unstructured text so downstream models and reviewers work from the material that actually matters.

  US 12,724,985 · granted September 1, 2026
• Sep 1 UnitedHealth

  Machine learning-based systems and methods for breath monitoring and assistance of a patient

  A wearable system that uses sensors and machine learning to track a patient's breathing in real time, detect splinting points where breathing is impeded, and deliver timed electrical stimulation to target nerves to help the patient breathe through those points.

  filed 2021 ▾

  A wearable breathing monitor carries one or more sensors that continuously observe a user's respiration, while machine-learning techniques determine when a full breath occurs and identify splinting points where breathing is impeded. When a splinting point is detected, a paired stimulator apparatus transmits stimulation signals timed to that moment, applying electrical pulses to target nerves according to a stimulation schedule. The closed loop of sensing, model-based breath analysis, and automatically controlled stimulation is intended to assist the patient's breathing in real time. The capability being protected is the coordination of continuous respiratory sensing with model-driven, precisely timed neurostimulation delivered through a wearable device.

  US 12,721,540 · granted September 1, 2026
• Sep 1 UnitedHealth

  Machine learning techniques for predictive schema analysis

  A predictive structural analysis system that applies machine learning models for table column classification, clustering, structural variance generation, and emergence reporting to infer and monitor the schema of tabular data.

  filed 2021 ▾

  The system performs predictive structural analysis on tabular data using a combination of machine-learning models: table column classification models label what each column represents, clustering models group related columns, structural variance generation models detect deviations in structure, and emergence report generation models produce descriptive output about newly appearing structures. Together these components let the system infer and monitor the schema of data tables rather than rely on fixed definitions. The approach suits environments where data layouts shift over time and must be reconciled automatically, such as ingesting varied healthcare or claims datasets. The filer is protecting a model-driven way to classify, cluster, and flag changes in data structure and to report on emerging schema patterns.

  US 12,724,751 · granted September 1, 2026
• Sep 1 Allstate

  Collision Analysis Platform Using Machine Learning to Reduce Generation of False Collision Outputs

  A telematics collision-detection platform that applies machine learning to sensor data and then filters false positives by checking whether the reading came from a known false-positive location and scoring the event against a telematics threshold, tightening the crash signals that feed auto claims.

  filed 2024 ▾

  The platform runs machine learning algorithms over received vehicle and telematics data to produce an initial collision output, then subjects any positive result to two suppression checks before affirming it. It first tests whether the data collection location falls within a set radius of a known false-positive location and, if so, reclassifies the event as a non-collision. Otherwise it computes a score from telematics data and compares it to a predetermined threshold, affirming a collision only when the score clears the bar. The output feeds first-notice-of-loss and automated claims triage for auto insurance, so the capability being protected is fewer spurious crash alerts and cleaner loss data.

  US 12,725,462 · granted September 1, 2026
• Sep 1 Humana

  Generating machine learning based models for time series forecasting

  A model-selection system for time series forecasting that evaluates a pool of machine learning models against a use-case-specific metric, selects the best performer, and uses it to forecast future values for that application.

  filed 2021 ▾

  For a given forecasting use case the system first determines a model metric appropriate to that use case, then draws on a pool of machine-learning models built from different techniques. It runs each candidate model to forecast the time series and computes the chosen metric for each, then selects the model whose metric value compares most favorably. That selected model is deployed to forecast future values of the time series for the application. The capability being protected is an automated, metric-driven way to match the right forecasting model to a particular use case, which in a health insurer's context could support projecting utilization, cost, or membership trends.

  US 12,725,056 · granted September 1, 2026
• Aug 25 UnitedHealth

  Systems and methods for utilizing topic models to weight mixture-of-experts for improvement of language modeling

  UnitedHealth's system tokenizes documents such as healthcare provider records, derives a topic vector, and weights multiple expert language models to combine their outputs into a next-text prediction.

  filed 2023 ▾

  The method receives documents, including documents associated with a healthcare provider, and processes them into tokens for downstream modeling. A topic model outputs a topic vector characterizing each document's subject matter, and the document is routed through multiple expert machine-learning models whose probability vectors are combined into a total probability vector weighted by that topic vector. A text output is selected using the combined vector, so the topic effectively controls how much influence each expert has on the final prediction. This topic-weighted mixture-of-experts design lets the language model specialize by subject matter, such as clinical or provider-record content, rather than treating all documents uniformly.

  What the examiner said

  Refused over one office action before it issued, on these grounds:

  • section 101, eligibility: whether this is patentable subject matter at all
  • section 103, obviousness: earlier references combined make it an obvious step

  The examiner called the claims directed to an abstract idea. It issued anyway, which is the ordinary outcome rather than a surprise.

  Cited against it: Lane et al.

  US 12,718,016 · granted August 25, 2026
• Aug 25 UnitedHealth

  Methods, systems, and computer program products for flagging dental claims for further scrutiny based on processing of dental clinical images and periodontal charts using multiple artificial intelligence (AI) models

  UnitedHealth's system flags dental claims for review by running clinical images through multiple AI models to identify procedure codes and using OCR on periodontal charts to read pocket measurements, then checking whether the submitted code is supported.

  filed 2023 ▾

  A clinical image tied to a dental procedure is processed through a plurality of AI models to independently identify one or more likely dental procedure codes, while a periodontal chart image is read via optical character recognition to obtain per-tooth pocket measurements and their positional coordinates. The system takes the procedure code submitted on the dental claim and first determines whether that procedure is visibly detectable in imagery. For visibly detectable procedures it checks whether the submitted code matches any AI-identified code, and it flags the claim when the codes do not match, when the procedure is not visibly detectable, or when the pocket measurements do not support the submitted code. This automates clinical review of dental claims for coding accuracy and potential fraud.

  What the examiner said

  Refused over one office action before it issued, on these grounds:

  • section 101, eligibility: whether this is patentable subject matter at all

  The examiner called the claims directed to an abstract idea. It issued anyway, which is the ordinary outcome rather than a surprise.

  US 12,718,300 · granted August 25, 2026
• Aug 25 State Farm

  Data mining system using artificial intelligence

  State Farm's document analysis system extracts embedded media and text from third-party filings, generates vector embeddings and a queryable dataset with metadata, and mines it to surface product modifications or new product ideas.

  filed 2024 ▾

  The system retrieves a document filed by a third-party entity along with its metadata, then separates embedded media into a media file and the text into a text file. It analyzes the media file to identify and extract data from within images or other embedded content, and processes the text into discrete text portions. For each text portion it generates embeddings, assembling a dataset that combines the text portions, embeddings, and metadata so the document can be queried for contextual data. The same dataset is then processed to identify potential product modifications or new products, turning unstructured third-party filings into structured, searchable signals for insurance product development.

  What the examiner said

  Refused over 2 office actions before it issued, on these grounds:

  • section 101, eligibility: whether this is patentable subject matter at all

  The examiner called the claims directed to an abstract idea. It issued anyway, which is the ordinary outcome rather than a surprise.

  US 12,717,800 · granted August 25, 2026
• Aug 25 State Farm

  Methods and apparatus to process insurance claims using cloud computing

  State Farm's cloud platform uses an AI resource to process information gathered from insureds or their representatives, identify the appropriate claim action, and route it for automated or manual claims processing.

  filed 2023 ▾

  The cloud platform includes an assistance resource that interacts electronically with an insured person or their representative to collect information related to a claim, serving as the intake point. An artificial intelligence resource processes the collected information to identify what action should be taken, effectively triaging the claim based on the gathered facts. A backend system carries out the identified action, and the same backend module remains usable by a human adjuster, so straightforward claims move through automated resolution while complex claims route to people using the same execution layer. The protected structure is this intake-to-AI-triage-to-shared-backend pipeline for cloud-based insurance claims processing.

  What the examiner said

  Refused over 4 office actions before it issued, on these grounds:

  • section 101, eligibility: whether this is patentable subject matter at all
  • section 103, obviousness: earlier references combined make it an obvious step

  US 12,718,299 · granted August 25, 2026

12 allowed8 final rejection11 responded38 under examination110 awaiting examination1 in examination

• State Farm awaiting examination published 9/3/2026

  Apparatuses, systems and methods for generating a base-line probable roof loss confidence score

  Describes a system that would generate a base-line probable roof loss confidence score from hail data, then use that score to produce property insurance underwriting, claims and loss mitigation data.

  How it works

  Combines a building's attributes, its roof characteristics and historical weather for its location into a base-line roof loss confidence score, then regenerates the score when a weather event occurs and, where the updated score deviates from the base-line beyond a set threshold, triggers a downstream action.

  Where it fits

  Property insurance underwriting, claims triage and loss mitigation, prioritizing roofs after hail and storm events.

  Stage

  Filed April 2026, published September 2026 and under examination; the claim read here is the original as filed, the broadest version and the one most likely to be narrowed.

  The claim as filed

  A computer-implemented method for property damage detection and mitigation comprising:

  1. receiving, by a processor of a computing device, building data representative of attributes of a building
  2. receiving, by the processor, roof data representative of a roof of the building
  3. receiving, by the processor, historical weather data for a geographic area that includes a location of the building
  4. generating, by the processor at a time for the building, a base-line probable roof loss confidence score based on the building data, the roof data, and the historical weather data
  5. responsive to receiving weather data indicating a weather event, generating an updated probable roof loss confidence score
  6. comparing, by the processor, the updated probable roof loss confidence score to the base-line probable roof loss confidence score generated at the time to determine a deviation
  7. implementing an action if the deviation exceeds a threshold. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260260139A1 · filed 4/22/2026
• Elevance awaiting examination published 8/20/2026

  System and Method for Generating High-Fidelity Privacy-Conscious Synthetic Patient Data for Causal Effect Estimation with Multiple Treatments

  Describes generating synthetic patient records modelled on a nationwide cohort, with the effects of treatments known in advance, so causal methods can be tested against a ground truth without exposing real patient data.

  How it works

  Filters a patient cohort to a treatment group, converts patient trajectories into standardized numeric tables, then trains a generator against a discriminator, weighting rarer features by an identifiability term and a Wasserstein-distance loss so the synthetic set tracks the real distribution while staying distinct from any real patient.

  Where it fits

  Health analytics and research, supplying shareable synthetic cohorts to study treatment effects without exposing real patient data.

  Stage

  Filed February 2026, published August 2026 and under examination; the claim read here is the original as filed, the broadest version and the one most likely to be rejected.

  The claim as filed

  A method of generating a synthetic patient dataset for use in causal effect evaluation of a treatment for a condition, from a real patient dataset comprising data relating to real patients who have been diagnosed with the condition, the method comprising the steps of:

  1. creating a filtered real patient dataset by removing, from the real patient dataset, data relating to patients having at least one of a disqualifying demographic condition, a disqualifying medical condition, and a disqualifying treatment
  2. filtering the real patient dataset to include only patients having received one of a plurality of identified treatments for the condition
  3. creating a standardized real patient dataset by capturing trajectory data representing the filtered real patient dataset
  4. converting the trajectory data to tabular data, the tabular data comprising samples relating to patients, and variables relating to patient features
  5. standardizing the variables into numerical values
  6. generating the synthetic patient dataset, using a generator, by creating a cartesian product of a patient feature space of the standardized real patient dataset, and a random variable space, and mapping the cartesian product to the patient feature space
  7. measuring a distance between a distribution of the synthetic patient dataset and a distribution of the patients in the standardized real patient dataset, using a discriminator, by mapping the patient feature space to a set of real numbers, wherein the generator and the discriminator are trained in an adversarial fashion
  8. applying an identifiability function to ensure that the synthetic patient dataset is substantially different from the standardized real patient dataset, wherein the identifiability function includes a weight calculated as an inverse of a discrete entropy of each patient feature, and wherein the weight of a patient feature has a direct correlation to a rarity of the patient feature
  9. applying a loss function to ensure that the distribution of the synthetic patient dataset is indistinguishable from the distribution of the patients in the standardized real patient dataset, wherein the loss function uses a Wasserstein distance between the distribution of the patients in the synthetic patient dataset and the distribution of the patients in the standardized real patient dataset, wherein the loss function includes a contrastive loss term, a

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260245743A1 · filed 2/9/2026
• State Farm awaiting examination published 8/20/2026

  Artificial intelligence-based query and response systems and methods

  Describes an AI system that drafts a proposed reply to an inbound customer message, routes it to a representative for feedback, and retrains itself on that feedback, aimed at customer servicing and contact handling.

  How it works

  Derives a query from an inbound message, sends a representative a model-generated proposed reply, records the representative's feedback together with that reply as a historical entry, compiles those entries into a training dataset, and retrains the AI model on it using machine learning.

  Where it fits

  Insurance customer service and claims correspondence, where representative feedback would continuously tune the automated replies.

  Stage

  Filed February 2025, published August 2026 and under examination; the claim read here is the original as filed, not a granted right.

  The claim as filed

  An AI-based computing system for responding in real-time to an inbound message, the computing system comprising:

  1. at least one memory for storing a plurality of category databases, wherein each category database includes one or more documents associated with the respective category of the category database
  2. at least one processor in communication with the at least one memory, the at least one processor configured to: transmit, to a representative computing device, an AImodel generated proposed response message responsive to a query message derived from the inbound message
  3. receive, from the representative computing device, feedback associated with the AI model generated proposed response message, the feedback provided by a representative
  4. create a historical record including the AImodel generated proposed response message and the feedback
  5. generate a training dataset including at least the created historical record
  6. using machine learning and/or artificial intelligence techniques, re-train the AImodel using the training dataset. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260244947A1 · filed 2/14/2025
• TIAA under examination published 8/20/2026

  Policy Definition and Enforcement Platform For AI-Based Access Control Systems

  Proposes deriving user and data-consumption profiles from information management data, then generating access controls from those profiles rather than from manually maintained permission lists.

  How it works

  Ingests information-management data, preprocesses it into user profiles and data-asset consumption profiles, feeds those to generative AI models trained on entitlement-labeled profiles to generate access controls, then checks each incoming request against the requesting user's assigned entitlements to allow or deny access.

  Where it fits

  Enterprise data governance and access control; the abstract describes general-purpose data security, not a specific insurance workflow.

  Stage

  Filed February 2025, published August 2026 and under examination; the claim read here is the original as filed and remains the applicant's request.

  The claim as filed

  A computer-implemented method for artificial intelligence-based data access control, the method comprising:

  1. obtaining, by one or more processors, information management data associated with a computing environment
  2. preprocessing, by the one or more processors, the information management data to generate (i) a set of user profiles corresponding to a plurality of users of the computing environment and (ii) data asset consumption profiles for the plurality of users
  3. generating, by the one or more processors, a set of access controls for the computing environment including a set of user data access entitlements by inputting the preprocessed information management data to one or more generative artificial intelligence (AI) models, wherein the one or more generative AI models are trained on user profiles labelled with corresponding user data access entitlements
  4. obtaining, by the one or more processors, a data access request from a user within the plurality of users
  5. identifying, by the one or more processors, a subset of the set of access controls including one or more user data access entitlements assigned to the user
  6. applying, by the one or more processors, the one or more user data access entitlements assigned to the user to the data access request to determine whether the data access request is valid or invalid
  7. in response to determining the data access request is valid, providing, by the one or more processors, data access to the user. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/31/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  The examiner has pointed at allowable subject matter in claims 3, 4, 5. That is the examiner naming the limitation that would earn the patent, not a statement that it will be granted.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260244775A1 · filed 2/19/2025
• State Farm responded published 8/20/2026

  Artificial intelligence-based query and response systems and methods

  Describes parsing an inbound message into keywords, retrieving documents from a matching subject-matter database, and passing them to a generative model that drafts a response and scores each document's relevance, aimed at customer contact handling.

  How it works

  Parses an inbound message into a query, extracts keywords to select a matching subject-matter document database, searches it for responsive documents and scores each for relevance, feeds those documents to a generative AI model to draft a reply, and notifies a representative with selectable document links each shown with its relevancy score.

  Where it fits

  Insurance contact-center triage, retrieving reference documents to draft and rank responses for a human representative.

  Stage

  Filed February 2025, published August 2026 and under examination; the claim read here has already been narrowed by amendment, not granted.

  The claim as amended, showing changes

  1. An Artificial Intelligence (Al) computer system for responding to an inbound message, the computer system comprising at least one processor, and at least one memory in communication with the at least one processor, the memory storing a plurality of subject-matter category databases, wherein each of the subject-matter category databases includes one or more documents associated with a subject-matter category, the at least one processor configured to: receive an inbound message from a requestor computing device
  2. parse the inbound message to generate a query
  3. determine one or more keywords contained within the query
  4. identify, based in part on the keywords, a subject-matter category database from the plurality of subject-matter category databases that is responsive to the one or more keywords
  5. search the identified subject-matter category database to determine one or more relevant documents, wherein the one or more relevant documents include information responsive to the query
  6. determine a relevancy score for each of the one or more relevant documents, the relevancy score indicating a level of relevance and responsiveness of the document to the inbound message
  7. input the determined one or more relevant documents into a generative Almodel to generate one or more model outputs including a proposed response message responding to the inbound message and the relevancy score for each of the relevant documents
  8. transmit a notification message to a representative computing device, wherein the notification message includes: i) the proposed response message, ii) a selectable link for one or more of the determined one or more relevant documents, and iii) [[a]] the relevancy score for each of the selectable links. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 3 office actions

  First action 10/1/2025, most recent 4/7/2026. The grounds raised so far:

  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  References cited against it: Smith, Gulwani et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260244670A1 · filed 2/14/2025
• State Farm allowed published 8/20/2026

  Artificial intelligence-based query and response systems and methods

  Describes a review interface that shows a representative an AI-drafted response with linked source documents, captures a response indicator, and retrains the model to display a revised reply, aimed at customer contact handling.

  How it works

  Displays to a representative an AI-drafted reply with document links and a feedback control, takes the assigned response indicator, routes the reply to a subject-matter expert for an updated indicator, records the exchange as a training record, retrains the response model, and shows a revised reply from the retrained model.

  Where it fits

  Insurance customer-service and claims desks, adding an expert-review step to the loop that tunes the automated replies.

  Stage

  Filed February 2025, published August 2026 and under examination; the claim read here was narrowed by amendment during prosecution, not allowed.

  The claim as amended, showing changes

  An AI-based computing system for responding in real- time to an inbound message, the AI-based computing system comprising:

  1. at least one memory for storing a plurality of subject-category databases, wherein each subject-category database includes one or more documents associated with [[the]] a respective subject-category of the subject-category database
  2. at least one processor in communication with the at least one memory, the at least one processor programmed to: cause to display, within a graphical user interface of a representative computing device, i) a proposed response message responsive to a query message derived from a current inbound message submitted by a user, the proposed response message being generated by an AIresponse model, ii) a link to one or more relevant documents retrieved from a corresponding subject-category database, and iii) a feedback input enabling a representative to assign a response indicator to the proposed response message
  3. receive, from the representative computing device, the response indicator assigned by the representative via the feedback input
  4. in response to receiving the response indicator, identify, using the AIresponse model, an expert computing device associated with a subject matter expert, wherein the subject matter expert is associated with a subject-category of the query message
  5. receive, from the expert computing device, an updated response indicator to the proposed response message, the updated response indicator assigned by the subject matter expert
  6. create a historical record including the proposed response message and the updated response indicator
  7. generate a training dataset including a plurality of historical records
  8. using machine learning and/or artificial intelligence tools, re-train the AIresponse model using the training dataset
  9. in response to a request for a revised response message, cause to display, within the graphical user interface of the representative computing device, in real-time, the revised response message responsive to the current inbound message using the re-trained AIresponse model. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 1/9/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Ponomarenko et al, Tomkins et al, Kanza et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260244626A1 · filed 2/14/2025
• Equifax awaiting examination published 8/13/2026

  Precursor learning for label refinement of machine learning training signals

  Proposes refining machine learning training labels by first training a classifier on a class-balanced older dataset, then using it to relabel misclassified rare-class cases in a newer dataset before the production model is trained.

  How it works

  Balances an older imbalanced dataset by pairing minority-class samples with an equal, randomly chosen set of majority-class samples, trains a classifier on it, then runs that classifier over a newer dataset to find majority-labelled cases that are really the rare class, relabels them, and trains the production model on the corrected data.

  Where it fits

  General-purpose model training; the claim describes rare-class detection in imbalanced data and names no specific insurance workflow.

  Stage

  Filed August 2025, published August 2026 and under examination; the claim read here is the original as filed, the broadest version and the one most likely to be narrowed.

  The claim as filed

  A system comprising:

  1. a non-transitory computer-readable medium comprising instructions that are executable by the processor to cause the processor to: receive a class-imbalanced example dataset comprising a multitude of individual majority class and minority class data samples previously collected during a first time interval
  2. combine a number of the minority class data samples with a randomly selected similar number of the majority class data samples to create a balanced training dataset
  3. initiate training of a data classification machine learning model using the balanced training dataset
  4. receive a class-imbalanced main dataset comprising a multitude of individual majority class and minority class data samples previously collected during a second time interval that is shorter and more recent than the first time interval
  5. execute the trained data classification machine learning model on the majority class data samples of the class-im balanced main data set to identify majority class data samples in the class-imbalanced main dataset that should be classified as minority class data samples
  6. create a refined dataset from the class-imbalanced main dataset by changing, from majority class to minority class, a classification of each data sample that is identified by the trained data classification machine learning model as being a minority class data sample
  7. create an updated main dataset by substituting the data samples of the refined dataset for the majority class data samples in the class-imbalanced main dataset
  8. initiate training of a main machine learning model using the updated main dataset. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260236835A1 · filed 8/14/2025
• Humana awaiting examination published 8/13/2026

  Ensemble time series model for forecasting

  Describes an ensemble forecaster that combines autoregressive features with embeddings drawn from sparse datasets, aimed at predictions that draw on several sources rather than a single series.

  How it works

  Takes two or more datasets of different categories, extracts a feature set from each, concatenates them into consolidated input vectors that combine observed historical values with fixed-length embeddings of time-lag sequences, then passes the vectors through a trained neural-network ensemble to generate prediction results.

  Where it fits

  General-purpose time-series forecasting; the abstract describes the model itself and names no specific insurance workflow.

  Stage

  Filed March 2026, published August 2026 and under examination; the claim read here is the original as filed, still the applicant's opening request.

  The claim as filed

  A computer-implemented method for generating time series predictions using an ensemble time series prediction model, the method comprising:

  1. receiving, by one or more processors, a set of data comprising two or more datasets of different categories
  2. extracting, by the one or more processors, a respective set of input features from each dataset of the two or more datasets
  3. forming, by the one or more processors, from the set of data, a set of consolidated input feature vectors by concatenating the respective sets of input features extracted from the two or more datasets, wherein the set of consolidated input feature vectors comprises a concatenation of at least: (i) historical observed values extracted from an observed historical dataset, and (ii) fixed-length embeddings of time-lag sequences generated from a time-lagged dataset comprising data with time lag information
  4. passing, by the one or more processors, the set of consolidated input feature vectors into the ensemble time series prediction model comprising a neural network
  5. generating, by the one or more processors, based on the set of consolidated input feature vectors, a set of prediction results using trained parameters of the ensemble time series prediction model. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260236743A1 · filed 3/27/2026
• State Farm awaiting examination published 8/13/2026

  Generating social media content for a user associated with an enterprise

  Seeks to fine-tune a base language model on enterprise data and use it to draft social media content shaped to a user profile, placing generative text inside a carrier's marketing workflow.

  How it works

  Claim 1 fine-tunes a base ML model on training data from many user profiles to produce a set of fine-tuned models, then on a request obtains the user's profile, loads the model tied to it, runs the request through a chatbot trained on that profile's communication style, and returns the generated social media content to the device.

  Where it fits

  Enterprise and agent marketing content generation, not a pricing, underwriting or claims function.

  Stage

  Filed April 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for generating social media content for a user associated with an enterprise using machine learning (ML), the computer-implemented method comprising:

  1. fine-tuning, by one or more processors, a base ML model based upon training data associated with a plurality of user profiles to generate a plurality of fine-tuned ML models
  2. receiving, by the one or more processors via a user device, a request for information
  3. obtaining, by the one or more processors, a user profile associated with a user
  4. providing, by the one or more processors, the request for the information to an ML chatbot based upon a fined-tuned ML model of the plurality of fine-tuned ML models, wherein: the ML chatbot is trained to generate a response based upon a user profile associated with the user
  5. the ML chatbot is trained using historical training data indicative of a style of communication associated with the user profile
  6. generating, by the one or more processors, the social media content based upon an output of the ML chatbot
  7. providing, by the one or more processors, the social media content to the user device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260236730A1 · filed 4/1/2026
• TIAA awaiting examination published 8/13/2026

  Artificial intelligence-based systems for generating personalized information

  Describes assembling a language model prompt from a user's identity and their activity during an authenticated session, so guidance is generated in context rather than drawn from fixed templates.

  How it works

  Claim 1 receives user input during an authorized session, builds a first LLM prompt carrying account identification and a request for additional identifying information, transmits it, receives that information back, then builds a second prompt combining the returned information with a request for personalized information tied to it.

  Where it fits

  Customer servicing and personalized member communications for a financial or insurance provider.

  Stage

  Filed February 2025, docketed and awaiting a first office action, on the original claims as filed, the applicant's broadest version.

  The claim as filed

  A computing system comprising:

  1. one or more processors
  2. one or more memories, having stored thereon instructions that, when executed, cause the computing system to: receive, via the one or more processors, an indication of user input associated with a user account during an authorized user session
  3. generate, via the one or more processors, a first language model prompt, the first language model prompt including (i.) identification information of the user account, and (ii.) a request for additional identifying information associated with the user account
  4. transmit, over a network interface, the first language model prompt
  5. receive, via the one or more processors, a response to the first language model prompt, the response including the additional identifying information
  6. generate, via the one or more processors, a second language model prompt, the second language model prompt including (i.) at least a portion of the additional identifying information received in response to the first language model prompt, and (ii.) a request for personalized information associated with the additional identifying information. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260236275A1 · filed 2/12/2025
• Allstate awaiting examination published 8/13/2026

  Processing System Having a Machine Learning Engine For Providing a Selectable Item Availability Output

  Describes a platform that updates availability determinations for selectable items as configuration and content change, with a machine learning engine producing the availability output.

  How it works

  Claim 1 stores historical availability inputs paired with result outputs, sorts them into per-result ML datasets, receives a new input set from a device, scores a match degree against each dataset, selects the dataset whose match exceeds a preset threshold, and returns that dataset's associated availability result to the device.

  Where it fits

  Availability or eligibility lookups presented to a customer, a servicing rather than pricing function.

  Stage

  Filed April 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computing platform, comprising:

  1. at least one processor
  2. a communication interface communicatively coupled to the at least one processor
  3. memory storing computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: store, in a machine learning database, a plurality of historical selectable item availability inputs, wherein each historical selectable item availability input is associated with a corresponding historical selectable item availability result output
  4. generate a plurality of machine learning datasets, wherein each machine learning dataset of the plurality of machine learning datasets corresponds to a different selectable item availability result output
  5. store the plurality of historical selectable item availability inputs in the plurality of machine learning datasets based on the corresponding historical selectable item availability result output associated with each historical selectable item availability input
  6. receive, from a user device and via a wireless data connection, a new set of selectable item availability inputs
  7. compare the new set of selectable item availability inputs to the plurality of machine learning datasets
  8. determine, for each machine learning dataset of the plurality of machine learning datasets, a match degree indicating a similarity between the new set of selectable item availability inputs and the historical selectable item availability inputs stored in that machine learning dataset
  9. identify a selected machine learning dataset from the plurality of machine learning datasets, wherein the selected machine learning dataset has a match degree that exceeds a predetermined threshold
  10. select a selectable item availability result output associated with the selected machine learning dataset
  11. send, to the user device and via the wireless data connection, the selected selectable item availability result output. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260236149A1 · filed 4/6/2026
• State Farm final rejection published 8/6/2026

  Artificial intelligence-based systems and methods utilizing smart building data analytics and loss reports

  Describes combining smart building sensor analytics with claims data in models trained per location to generate loss-related recommendations for a chosen site, aimed at property risk assessment and loss prevention.

  How it works

  Claim 1 as amended receives smart building analytics including electrical sensor data from many buildings plus claims data from an overlapping set, takes a target location, runs both through AI models trained for that location, and outputs building-plan recommendations drawn from buildings whose characteristics are similar to the target.

  Where it fits

  Property loss mitigation and pre-loss risk assessment for insured structures.

  Stage

  Filed May 2025 and under a final rejection: the examiner has rejected the current claims, and this claim was already narrowed during prosecution rather than agreed.

  The claim as amended, showing changes

  1. A building planning computer system for generating a building plan using an artificial intelligence (Al) model component, the building planning computer system comprising at least one processor, an Almodel component comprising one or more Almodels, and at least one memory device, wherein the at least one processor is programmed to: receive smart building analytics data associated with a first plurality of buildings each located at different locations, wherein the smart building analytics data includes at least electrical sensor data from each of the first plurality of buildings
  2. receive claims data associated with a second plurality of buildings each located at different locations, wherein the first plurality of buildings includes at least some of the second plurality of buildings
  3. receive input data including at least a select location
  4. access the one or more artificial intelligence (Al) models trained to analyze input data associated with the select location
  5. input the smart building analytics data and the claims data into the one or moreAlmodels to generate one or more outputs including recommendations for a building plan associated with the select location based upon the smart building analytics data and the claims data from buildings at locations having similar characteristics to the select location
  6. and transmit the one or morerecommendations to a user computing device.

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 4/3/2026, most recent 8/19/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260228840A1 · filed 5/14/2025
• State Farm responded published 8/6/2026

  Artificial intelligence-based systems and methods utilizing smart building data analytics and loss reports

  Describes training models on building sensor analytics and claims data to generate a construction plan, with a materials list and drawings, for a building designed to reduce the likelihood of loss at a chosen location.

  How it works

  Claim 1 as amended takes analytics describing each building's materials, features and surroundings plus claims loss data with the event behind each loss, trains AI models to output a plan whose recommended materials and features are customized to reduce loss at a target location, then outputs a materials list and design drawings.

  Where it fits

  Property loss mitigation through risk-informed construction guidance for insured structures.

  Stage

  Filed May 2025, the applicant having responded to an office action, on a claim already narrowed during prosecution rather than the opening version.

  The claim as amended, showing changes

  1. A building planning computer system for generating a building plan using an artificial intelligence (Al) model component, the building planning computer system comprising at least one processor, an Almodel component comprising one or more Almodels, and at least one memory device, wherein the at least one processor is programmed to: receive smart building analytics data associated with a first plurality of buildings each located at different locations, wherein the smart building analytics data includes at least data describin g: (i ) materials used to build each respective building, (ii) building features included within each respective building, and (iii) features surrounding each respective building of the first plurality of buildings
  2. receive claims data associated with a second plurality of buildings each located at different locations, wherein the first plurality of buildings includes at least some of the second plurality of buildings, and wherein the claims data includes loss data associated with each respective building and an event associated with each loss
  3. train the one or more Almodels using the smart building analytics data and the claims data, the one or more Almodels trained to output a building plan for constructing an enhanced building at a select location, wherein the enhanced building includes recommended materials orand/irecommended building featuresand/or recommended surrounding features customized for the select location that reduce an overall likelihood of loss at the enhanced building
  4. input into the one or more Almodels construction data for constructing the enhanced building at the select location
  5. output the building plan for the enhanced building including a recommended materials list and design drawings including areas surrounding the enhanced buildingfor constructing the enhanced building based upon the construction data. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 4/20/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260228839A1 · filed 5/14/2025
• State Farm awaiting examination published 8/6/2026

  Augmented reality system to provide recommendation to purchase a device that will improve home score

  Proposes overlaying recommended devices onto a live camera view of a property by correlating underlay and overlay image layers, so a homeowner sees suggested equipment in place before buying it.

  How it works

  Claim 1 receives underlay layer data for an AR viewer's field of view, calculates the improvement to a home score from placing a recommended device, receives overlay data naming the device and the score improvement, correlates overlay with underlay, and builds an AR display showing the device and the score gain to the user.

  Where it fits

  Home risk-mitigation guidance and loss-prevention engagement for property policyholders.

  Stage

  Filed April 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method of using Augmented Reality (AR) for visualizing a recommended device for a home, the computer-implemented method comprising:

  1. receiving, with one or more processors, underlay layer data indicative of a field of view associated with an AR viewer device
  2. calculating, with the one or more processors, an improvement to a home score associated with the home based upon placement of the recommended device for the home
  3. receiving, with the one or more processors, overlay layer data including an indication of the recommended device and the improvement to the home score
  4. correlating, with the one or more processors, the overlay layer data with the underlay layer data
  5. creating, with the one or more processors, an AR display based upon the correlation, the AR display including an illustration of the recommended device for the home and the improvement to the home score
  6. displaying, with the one or more processors, the AR display to a user via the AR viewer device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260228831A1 · filed 4/2/2026
• State Farm awaiting examination published 8/6/2026

  Artificial intelligence-based systems and methods utilizing smart building data analytics and loss reports

  Seeks to train a model on smart building analytics and claims data to generate construction plans for a subdivision whose materials and features are chosen to lower expected loss and improve energy efficiency.

  How it works

  Claim 1 takes building analytics from many buildings plus claims data from an overlapping set, trains AI models to output a plan for a subdivision of enhanced buildings whose materials and features are customized to reduce loss and whose positioning improves energy efficiency, then outputs a materials list and design drawings.

  Where it fits

  Property loss mitigation and pre-loss risk assessment at a subdivision scale.

  Stage

  Filed May 2025, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  1. A building planning computer system for generating a building plan using an artificial intelligence (AI) model component, the building planning computer system comprising at least one processor, an AI model component comprising one or more AI models, and at least one memory device, wherein the at least one processor is programmed to: receive smart building analytics data associated with a first plurality of buildings each located at different locations
  2. receive claims data associated with a second plurality of buildings each located at different locations, wherein the first plurality of buildings includes at least some of the second plurality of buildings
  3. train the one or more AI models using the smart building analytics data and the claims data, the one or more AI models trained to output a building plan for constructing an enhanced subdivision of a plurality of enhanced buildings at a select location, wherein each enhanced building includes materials and/or features customized for the select location that reduce an overall likelihood of loss at the corresponding enhanced building, and wherein the enhanced subdivision includes features and/or building positioning that improve an overall energy efficiency of the plurality of enhanced buildings
  4. input construction data into the one or more AI models for constructing the enhanced subdivision at the select location
  5. output the building plan for the enhanced subdivision and each of the enhanced buildings including a materials list and design drawings. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260228372A1 · filed 5/14/2025
• TIAA awaiting examination published 7/30/2026

  Systems and methods for artificial intelligence - blockchain retirement account management

  Describes retrieving retirement account data from third-party institutions and passing it through a trained capture model, aimed at discovering and acting on accounts a saver may have lost track of.

  How it works

  Claim 1 takes a transaction request with retirement account identifiers, retrieves plan rules or regulations, feeds the request and rules to a trained smart-contract model to generate contract code, compiles and deploys it to a distributed ledger, sends inputs to the contract, and returns the outputs as a summary to the user.

  Where it fits

  Retirement account transaction execution and recordkeeping for a plan provider.

  Stage

  Filed March 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for performing retirement account transactions on a distributed ledger, the method comprising:

  1. receiving, from a user by one or more processors, a transaction request comprising a transaction description and one or more retirement account identifiers associated with one or more retirement accounts
  2. retrieving, by the one or more processors, plan requirements associated with the one or more retirement accounts, wherein the plan requirements comprise financial institution rules or government regulations
  3. providing, by the one or more processors, the transaction description, the one or more retirement account identifiers, and the plan requirements to a trained smart contract model to cause the trained smart contract model to generate smart contract code, wherein the smart contract code comprises programmatic logic for executing the transaction request
  4. compiling, by the one or more processors, the smart contract code into a smart contract, wherein the smart contract is configured to: receive one or more transaction inputs associated with the transaction request, execute the transaction request using the one or more transaction inputs, and generate one or more transaction outputs
  5. transmitting, by the one or more processors, the smart contract to at least one other participant in a distributed ledger network to deploy the smart contract, wherein the smart contract is stored in the distributed ledger maintained by a network of participants
  6. sending, by the one or more processors, the one or more transaction inputs to the smart contract
  7. receiving, by the one or more processors, the one or more transaction outputs from the smart contract
  8. providing, by the one or more processors to the user, a transaction summary comprising information about the one or more transaction inputs and the one or more transaction outputs. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260220711A1 · filed 3/20/2026
• TIAA under examination published 7/30/2026

  Systems and methods for artificial intelligence - blockchain retirement account management

  Describes retrieving retirement account data from third-party institutions, structuring it with a capture model and forecasting with a deep learning model, aimed at discovering and managing scattered accounts.

  How it works

  Claim 1 authenticates the user by decrypting credentials with a public key, receives a request to find lost retirement accounts, retrieves the user's identification data, supplies it to several third-party institutions to search for matching accounts, then on a hit provides reset links and saves a returned account identifier and credentials.

  Where it fits

  Account aggregation and onboarding for retirement savers, a servicing function.

  Stage

  Filed March 2026 and under examination, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for discovering lost retirement accounts for a user, the method comprising:

  1. receiving, from the user by one or more processors, user authentication credentials comprising data encrypted by a private key associated with the user
  2. authenticating, by the one or more processors, the user by decrypting the data by a public key associated with the private key
  3. receiving, from the user by the one or more processors, a request to discover the lost retirement accounts
  4. retrieving, by the one or more processors, identification data associated with the user
  5. supplying, by the one or more processors, the identification data to a plurality of third-party financial institutions to cause the plurality of third-party financial institutions to search for one or more retirement accounts associated with the identification data
  6. responsive to receiving, by the one or more processors, notification of the one or more retirement accounts associated with the identification data: providing, by the one or more processors, one or more links to one or more of the plurality of third-party financial institutions associated with the one or more retirement accounts to enable the user to reset third-party authentication credentials, receiving, by the one or more processors, an account identifier and the third-party authentication credentials for the one or more retirement accounts, and saving, by the one or more processors, the account identifier and the third-party authentication credentials. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/11/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260220710A1 · filed 3/20/2026
• Equifax awaiting examination published 7/30/2026

  Explainable machine-learning techniques from multiple data sources

  Describes a hybrid machine learning model for risk assessment that would train on a loss function combining discriminative and generative terms to handle missing input values, and output both a risk indicator and the explanatory data behind it.

  How it works

  The model would compute a risk indicator for a target entity from its predictor variables, training by iteratively minimizing a loss function that blends a discriminative and a generative term so it fits even when some input values are unknown, then generate explanatory data linking changes in the score to changes in specific predictors.

  Where it fits

  Automated risk assessment and scoring that gates access to interactive computing services; general risk scoring rather than a named insurance line.

  Stage

  Filed June 2025 and published July 2026; under examination on the claim as originally filed, which has not been narrowed in prosecution yet.

  The claim as filed

  A method performed by one or more processing devices, comprising:

  1. determining, using a hybrid machine learning model trained using a training process, a risk indicator for a target entity from predictor variables associated with the target entity, wherein the risk indicator indicates a level of risk associated with the target entity, wherein the training process includes operations comprising: accessing training vectors having a plurality of sets of training predictor variables and a plurality of training outputs corresponding to the respective sets of training predictor variables
  2. performing iterative adjustments of parameters of the hybrid machine learning model to minimize a loss function of the hybrid machine learning model, wherein the loss function comprises a first term representing a discriminative loss and a second term representing a generative loss, wherein a value of a predictor variable in the predictor variables associated with the target entity is unknown or a value of a training predictor variable or a training output in the training vectors is unknown
  3. generating, for the target entity, explanatory data indicating relationships between changes in the risk indicator and changes in at least some of the predictor variables associated with the target entity
  4. transmitting, to a remote computing device, a responsive message including at least the risk indicator and the explanatory data for use in controlling access of the target entity to one or more interactive computing environments. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260220531A1 · filed 6/20/2025
• AIG allowed published 7/30/2026

  User interface for providing source traceability within an information extraction system

  Describes an information extraction system that tags document chunks with identifiers and asks a language model to cite which chunks it relied on, so each extracted value can be traced, verified, and audited, with an interface for human review and correction.

  How it works

  Splits documents into chunks, each tagged with an identifier pointing back to its source, searches for chunks relevant to a query, prompts a language model to extract the target information and to name which chunks it used, then builds an interface showing each extracted value beside a citation to its source passage.

  Where it fits

  General-purpose extraction from unstructured documents; the abstract describes source traceability and audit rather than a named insurance workflow.

  Stage

  Filed January 2026, published July 2026 and allowed after the applicant narrowed the claim during prosecution; a notice of allowance is not an issued patent.

  The claim as amended, showing changes

  A method for providing source content for information extracted by language models, the method comprising:

  1. generating, by one or more processors, a plurality of chunks from one or more documents, each chunk associated with an identifier referring to a portion of the one or more documents that corresponds to the chunk
  2. identifying, by the one or more processors, one or more relevant chunks from the plurality of chunks based on a search criterion
  3. transmitting, by the one or more processors to a language model, a prompt comprising (i) a first request to extract particular information from the one or more relevant chunks and (ii) a second request for the language model to identify one or more used chunks of the one or more relevant chunks used to extract the particular information
  4. receiving, by the one or more processors, a response from the language model comprising the particular information extracted and identifiers of the one or more used chunks identified by the language model
  5. and generating, by the one or more processors, instructions for a user interface comprising the particular information extracted by the language model and a citation to the portion of the one or more documents from which the particular information is extracted, based on the identifiers of the one or more used chunks received in the response from the language model. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 5/14/2026. The grounds raised so far:

  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Veillon et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260220372A1 · filed 1/12/2026
• AIG allowed published 7/30/2026

  Information extraction system for unstructured documents using independent tabular and textual retrieval augmentation

  Describes a retrieval system that separates a document's tables from its text, indexes each as vector embeddings, and sends only the chunks relevant to a query to a language model, so tabular and prose content are retrieved independently when data elements are extracted.

  How it works

  Receives content containing text and tables, separates them into table chunks and text chunks, converts each chunk into a vector embedding with an index entry, finds the table or text chunk relevant to a prompt through a search of those entries, and stores the language model's response together with the chunk it drew on.

  Where it fits

  General document extraction where tabular data must be retrieved separately from prose; the abstract names no specific insurance workflow.

  Stage

  Filed September 2025, published July 2026 and allowed after the applicant narrowed the claim during prosecution; the issue fee is paid but the patent has not yet issued.

  The claim as amended, showing changes

  A method for prompting a language model with content, the method comprising:

  1. receiving, by one or more processors, the content including text and one or more tables
  2. forming, by the one or more processors, one or more table chunks and one or more text chunks from the content by separating the one or more tables from the text, wherein each table chunk of the one or more table chunks comprising at least a portion of a table of the one or more tables and each text chunk comprising at least a portion of the text
  3. generating, by the one or more processors, one or more first index entries for the one or more table chunks and one or more second index entries for the one or more text chunks by converting the one or more table chunks and the one or more text chunks into vector text embeddingsusing a text embedding model : identifying, by the one or more processors, a relevant table chunk of the one or more table chunks or a relevant text chunk of the one or more text chunks based on a search criterionsearch of the one or more first index entries or the one or more second index entries related to a prompt for the language model
  4. storing a response from the language model to the prompt with the relevant table chunk or the relevant text chunk. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 2/19/2026, most recent 4/14/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Vellon et al, Thompson et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260220371A1 · filed 9/12/2025
• Allstate awaiting examination published 7/23/2026

  Applying Machine Learning to Telematics Data to Predict Accident Outcomes

  Seeks to predict claim outcomes from telematics and related data, including which parts of a vehicle were damaged and the expected liability, moving loss estimation ahead of first notice of loss.

  How it works

  Claim 1 obtains accelerometer and GPS data sampled at different rates for a crashed vehicle, resamples both to a common rate, computes vectors and acceleration signals, enriches with external context such as intersection, weather or speed limit, and feeds a classifier trained on historical telematics and outcomes to predict the outcome.

  Where it fits

  Auto claims triage and outcome prediction from post-accident telematics.

  Stage

  Filed November 2025, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A method comprising:

  1. obtaining, by a computing device, telematics data for a vehicle involved in a vehicular accident, the telematics data comprising acceleration data sampled at a first sampling rate using an accelerometer and location data sampled at a second sampling rate using a GPS receiver, wherein the first sampling rate is different from the second sampling rate
  2. converting, by the computing device, the acceleration data and the location data to a common sampling rate by down-sampling the acceleration data and up-sampling the location data
  3. computing, by the computing device, vectors based on the acceleration data at the common sampling rate and heading data
  4. calculating, by the computing device, acceleration signals based on the computed vectors
  5. enriching, by the computing device, the telematics data by incorporating data from an external database to generate enriched telematics data, wherein the enriching comprises at least one of : determining whether the vehicle was in an intersection based on GPS coordinates, determining weather conditions at a location of the vehicle at a time of the vehicular accident, or determining a speed limit on a road the vehicle was traveling at the time of the vehicular accident
  6. inputting, by the computing device, the enriched telematics data into a machine learning classifier trained on historical telematics data and historical accident outcome data to obtain an accident prediction for the vehicle. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260212422A1 · filed 11/10/2025
• Equifax awaiting examination published 7/23/2026

  Machine-learning for content interaction

  Describes building linked graph structures from entity and interaction data, then selecting operations for a model to generate a content recommendation that facilitates an interaction between parties.

  How it works

  Claim 1 takes a provider's request to recommend content for an interaction with a target entity, builds two graph structures from the entity's data and interaction data, links them, chooses which of several operations to run on the linked graph, executes them with a trained model to generate a content recommendation, and returns a message.

  Where it fits

  Customer interaction and content targeting; the abstract states no specific insurance function.

  Stage

  Filed June 2025, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A system comprising:

  1. a non-transitory computer-readable medium comprising instructions that are executable by the processor to cause the processor to perform operations comprising: receiving a request from a provider entity, the request including a request to provide content recommendation to facilitate an interaction between the provider entity and a target entity
  2. receiving entity data and interaction data associated with the target entity
  3. generating, based on the entity data and the interaction data, at least a first graph structure and a second graph structure
  4. generating, based on the first graph structure and the second graph structure, a linked graph structure
  5. determining, among a plurality of operations, one or more target operations to perform on data included in the linked graph structure
  6. executing, using a trained machine-learning model, the one or more target operations on the linked graph structure to generate a content recommendation to facilitate the interaction
  7. providing a responsive message based on the content recommendation usable to facilitate the interaction. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260212263A1 · filed 6/25/2025
• State Farm awaiting examination published 7/16/2026

  Sensing peripheral heuristic evidence, reinforcement, and engagement system

  Describes analysing household sensor data, including device-level electricity use, to identify a condition affecting a resident, aimed at detecting change in a home without direct observation.

  How it works

  Claim 1 analyzes data from home-mounted sensors with a model trained on that home's historical sensor and condition data, determines an abnormal condition of an individual, including detecting an atypical movement pattern from the sensor data, and generates an electronic notification of that condition.

  Where it fits

  Home-based risk monitoring and caregiver alerting, adjacent to health and life risk assessment.

  Stage

  Filed March 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for analyzing sensor data captured by one or more home-mounted sensors associated with a home environment using a trained machine learning model to identify an abnormal condition of an individual in the home environment, the method comprising:

  1. analyzing sensor data captured by one or more home-mounted sensors associated with a home environment using a trained machine learning model, the trained machine learning model being trained using historical sensor data collected the home environment and historical condition data, the historical condition data indicating conditions associated with at least one individual in the home environment
  2. determining an abnormal condition of an individual in the home environment based upon the analysis using the trained machine learning model, wherein the determining the abnormal condition of the individual in the home environment comprises determining an atypical movement pattern using the trained machine learning model based upon the sensor data
  3. generating an electronic notification indicating the abnormal condition of the individual via one or more processors. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260204149A1 · filed 3/9/2026
• State Farm awaiting examination published 7/16/2026

  Intelligent machine-learned model monitoring

  Proposes inferring the accuracy and drift of third-party machine learned models from their inputs and outputs, then switching or retraining downstream models when a monitored model degrades.

  How it works

  Claim 1 sends the same unstructured input to an obscured third-party processing system at two times, receives labeled output each time, computes a difference between the two label sets, tests it against a threshold, and when it is met initiates a reconfiguration of a downstream system that consumes the third-party output.

  Where it fits

  Model risk governance and monitoring of vendor models used in insurance analytics.

  Stage

  Filed January 2025, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for inferring obscured data processing system performance, the computer-implemented method comprising:

  1. transmitting, by a processor at a first time, first unstructured input data to an obscured data processing system
  2. receiving, at the processor from the obscured data processing system, first labeled output data associated with the first unstructured input data
  3. transmitting, by the processor at a second time subsequent to the first time, the first unstructured input data to the obscured data processing system
  4. receiving, at the processor from the obscured data processing system, second labeled output data associated with the first unstructured input data
  5. determining, at the processor, based on first labels represented in the first labeled output data and second labels represented in the second labeled output data, a difference value indicating a difference between the first labeled output data and the second labeled output data
  6. determining, at the processor, that the difference value meets or exceeds a difference threshold
  7. initiating, at the processor, based on determining that the difference value meets or exceeds a difference threshold, a reconfiguration of a computing system configured to consume output generated by the obscured data processing system. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260203648A1 · filed 1/15/2025
• Equifax awaiting examination published 7/16/2026

  Machine-learning techniques for risk assessment based on clustering

  Describes matching an entity's historical risk data to a cluster derived from high-dimensional clustering across many entities, then predicting future risk from that cluster rather than the entity alone.

  How it works

  Claim 1 identifies, from clusters built by high-dimensional clustering over many entities, the target cluster matching an entity's historical risk data, finds that cluster's nearest neighboring clusters, and determines a prediction of future risk for the entity from the target cluster and its neighbors.

  Where it fits

  Risk scoring and access control, applicable to underwriting-style risk assessment.

  Stage

  Filed March 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A method comprising:

  1. identifying, by a processing device, a target cluster out of a plurality of clusters, the target cluster matching historical risk assessment data of a target entity, wherein the plurality of clusters are determined based on risk assessment data of a plurality of entities using high dimensional clustering
  2. identifying, by the processing device and from the plurality of clusters, a set of nearest neighboring clusters of the target cluster
  3. determining, by the processing device, a prediction of risk for the target entity based on the target cluster and the set of nearest neighboring clusters. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260203611A1 · filed 3/11/2026
• State Farm awaiting examination published 7/16/2026

  Structured data extraction using generative machine learning models

  Seeks to segment unstructured text, categorise each segment, then build a prompt per segment so a generative model returns structured fields, aimed at extraction and routing without fixed parsers.

  How it works

  Claim 1 takes a file of customer-interaction text, assigns categories to different portions, derives a field name and description for a portion, builds a prompt from them, passes the prompt and text to an ML model to extract a value, checks it against a constraint, and on satisfaction stores it as metadata linked to the file.

  Where it fits

  Claims and servicing intake, turning unstructured customer text into validated structured fields.

  Stage

  Filed March 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method, comprising:

  1. receiving, by a processor, an electronic file including text indicative of a customer interaction
  2. determining, by the processor, a first category associated with a first portion of the text and a second category associated with a second portion of the text
  3. determining, by the processor and based at least in part on the first category, a first field name of a first data field, and a first description of the first data field
  4. generating, by the processor, a prompt based at least in part on the first field name and the first description of the first data field
  5. providing, by the processor, the prompt and the first portion of the text to a machine learning model
  6. determining, by the processor, based at least in part on the machine learning model, an output identifying a first value corresponding to the first data field
  7. determining, by the processor, that the first value satisfies a first constraint
  8. based on determining that the first value satisfies the first constraint, generating, by the processor, first metadata corresponding to the electronic file, the first metadata representing the first value for the first data field
  9. storing, by the processor, the first metadata in a first database and in association with the electronic file. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260203341A1 · filed 3/9/2026
• Allstate awaiting examination published 7/9/2026

  Systems and Methods For Predicted Total Loss Determinations Based On Image Analysis

  Proposes analysing uploaded and labelled images to identify a property type, match it against the reported property, and predict whether a loss should be treated as a total loss.

  How it works

  Claim 1 takes labeled images and a claim identifier, retrieves the claim's total loss score and reported property type, uses an AI model to separate usable from unusable images, identifies the property type from the usable ones, matches it to the reported type, and generates a predicted total loss determination from the score, images and match.

  Where it fits

  Auto or property claims total-loss adjudication from submitted images.

  Stage

  Filed October 2025 and awaiting examination, on a new claim presented after the original claims were cancelled, typically following a rejection.

  The claim new claim

  An intelligent prediction system comprising:

  1. one or more processors
  2. one or more memory components communicatively coupled to the one or more processors
  3. machine readable instructions stored in the one or more memory components that cause the intelligent prediction system to perform at least the following when executed by the one or more processors: receive from a user one or more uploaded and labeled images of at least a property and a claim identifier associated with the property
  4. retrieve information based on the claim identifier, wherein the information comprises a total loss score of the property and a reported property type of the property
  5. filter, via an artificial intelligence model of a data analytics module, the uploaded and labeled images to determine one or more usable filtered images and one or more unusable images
  6. analyze the one or more usable filtered images of the one or more uploaded and labeled images to generate one or more processed images based on the one or more usable filtered images via the artificial intelligence model of the data analytics module such that the one or more processed images exclude the one or more unusable images
  7. determine an identified property type of the property from the one or more processed images
  8. generate a property match indication between the identified property type determined from the one or more processed images generated based on the one or more usable filtered images and the reported property type associated with the claim identifier when there is a match therebetween
  9. generate an automated predicted total loss determination based on the total loss score, the one or more processed images from the data analytics module generated based on the one or more usable filtered images, and the property match indication. Docket No.: 20665-2243/AL T000 9NA New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260195820A1 · filed 10/13/2025
• State Farm awaiting examination published 7/9/2026

  Media enhancement virtual assistant

  Describes scoring agent social media profiles with a media enhancement model, aimed at ranking and improving how individual agents present a carrier's brand online.

  How it works

  Claim 1 scores each agent profile with a trained model that weights follower change in the social media data, partitions the scores into agent groups by a statistical distribution, derives top posts and a per-agent top-posts list by semantic analysis, generates curated tasks to improve the score, and updates them as engagement changes.

  Where it fits

  Agent marketing and social media engagement support, not a pricing or underwriting function.

  Stage

  Filed February 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A virtual assistant device for enhancing media content and presence, comprising:

  1. one or more processors
  2. a memory communicatively coupled to the one or more processors, the memory containing instructions therein that, when executed, cause the one or more processors to: determine a composite score corresponding to each respective agent profile of a plurality of agent profiles by applying an enhancement model to social media data from the plurality of agent profiles, wherein the enhancement model is a trained machine learning model that evaluates a change in followers included in the social media data for a respective agent profile in accordance with a weighting algorithm to determine the composite score corresponding to each respective agent profile
  3. partition, by the enhancement model, the composite scores to correspond with one or more agent profile groups based upon satisfying a statistical distribution of the one or more agent profile groups
  4. determine (i) one or more top posts by applying the enhancement model to the one or more agent profile groups and the social media data and (ii) a top posts list for a respective agent based upon a semantic analysis of posts in a respective posting history associated with the respective agent profile
  5. generate, by the enhancement model, one or more curated tasks indicating one or more current trends configured to improve the composite score of the respective agent profile based upon the top posts list, a respective agent profile group corresponding to the respective agent profile, and the social media data
  6. determine, by the enhancement model, an updated composite score based upon completion of at least a portion of the one or more curated tasks
  7. generate, by the enhancement model, updated curated tasks based upon a change in engagement after at least the portion of the one or more curated tasks have been completed
  8. display at least one of the top posts list, the one or more curated tasks, or the updated curated tasks for viewing by a respective agent associated with the respective agent profile. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260195721A1 · filed 2/26/2026
• TIAA responded published 7/9/2026

  Techniques for Self-Guided Hyper Personalization Governance Using Nested Machine Learning Models

  Describes governing operating parameters through nested machine learning models built from multiple agents, so personalisation settings are adjusted by the system rather than configured by hand.

  How it works

  Claim 1 as amended receives knowledge data indicating operating parameters, analyzes an input tied to a subset of them with a nested model of several agent models trained on composite knowledge data, determines a compliance action to configure that subset, and generates a recommendation tied to the compliance action.

  Where it fits

  Governance and compliance workflow support for a financial or insurance provider.

  Stage

  Filed January 2025, the applicant having responded to an office action, on a claim already narrowed during prosecution rather than the opening version.

  The claim as amended, showing changes

  A computer-implemented method for managing governance operating parameters using machine learning, the computer-implemented method comprising:

  1. receiving, by one or more processors, knowledge data, wherein the knowledge data is indicative of one or more operating parameters
  2. analyzing, by the one or more processors using a nested machine learning model comprising a plurality of agent models trained using composite knowledge data, an input associated with a subset of the one or more operating parameters indicated by the composite knowledge data, wherein the composite knowledge data is generated by one or more data processing machine learning models based on the knowledge data
  3. determining, by the one or more processors based on the analyzing, a compliance action, wherein the compliance action is associated with configuring the subset of the one or more operating parameters associated with the input
  4. generating, by the one or more processors using the nested machine learning model, a recommendation associated with the compliance action. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 5/13/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260195695A1 · filed 1/3/2025
• State Farm awaiting examination published 7/9/2026

  System and methods for predictive modeling based upon multimodal geotagged data

  Seeks to combine geotagged data from user devices with supplemental location data in a model that predicts events and builds a location risk profile from their frequency, aimed at location risk assessment.

  How it works

  Claim 1 takes geotagged data from a user's devices, gathers supplemental data for the tagged locations, feeds both into a model trained to predict event occurrences and recommendations at those locations, builds a risk profile from the frequency of predicted events, and outputs a prediction presentation to the user's interfaces.

  Where it fits

  Location-based risk assessment and loss-prevention alerts for policyholders.

  Stage

  Filed January 2025, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for utilizing geotagged data for predictive modeling, the computer-implemented method performed by one or more processors of a computing system in communication with one or more data sources, the computer-implemented method comprising:

  1. receiving, by the one or more processors, a first set of geotagged data from one or more devices associated with a user
  2. processing, by the one or more processors, data received from the one or more data sources for supplemental data corresponding to one or more locations in the first set of geotagged data
  3. inputting, by the one or more processors, the first set of geotagged data and the supplemental data into a machine-learning model, wherein the machine-learning model is trained to generate (i) an event prediction corresponding to one or more event occurrences at the one or more locations, and (ii) one or more recommendations corresponding to the one or more predicted events
  4. generating, by the one or more processors, a risk profile for the one or more locations based upon a frequency of the one or more event occurrences of the one or more predicted events
  5. outputting, by the one or more processors, a prediction presentation based upon the event prediction, the risk profile, and the one or more recommendations to one or more user interfaces of the one or more devices associated with the user. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260195641A1 · filed 1/28/2025
• LexisNexis Risk awaiting examination published 7/9/2026

  Systems and methods for chatbot authentication

  Describes an authentication system that generates open-ended questions from known data about a user and uses a large language model to judge whether natural-language answers are correct, incorrect, or incomplete, asking follow-ups before granting access.

  How it works

  Resolves supplied user information to a unique identifier, pulls matching data from data sources, generates authentication questions from it, takes the user's free-text answers, and uses a language model to score each answer against the known ground truth as correct, incorrect, or incomplete before authenticating or asking for more detail.

  Where it fits

  Identity verification and account-access authentication, a control that supports fraud prevention.

  Stage

  Filed March 2026, published July 2026 and awaiting examination; the claim read here is the original as filed, the broadest version and the one most likely to be rejected.

  The claim as filed

  A computer-implemented method for user authentication for access to a service using a chatbot, the method comprising:

  1. receiving user information corresponding to a user
  2. resolving the received user information to a unique identifier (UID) for the user
  3. obtaining, from one or more data sources, data about the user that matches the UID
  4. generating a plurality of authentication questions based on the data
  5. outputting for display on a user device associated with the user, one or more of the plurality of authentication questions
  6. receiving, in a natural language, one or more user answers corresponding to the one or more of the plurality of authentication questions
  7. utilizing a large language model (LLM) to evaluate the one or more user answers against a known ground truth based on the data, wherein the evaluation includes determining if the one or more user answers are factually correct, incorrect, or incomplete
  8. responsive to the evaluation, authenticating the user for access to the service or providing iterative feedback to the user. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260195431A1 · filed 3/3/2026
• Cigna responded published 7/2/2026

  Contact center intelligent data retrieval system and method

  Seeks to route contact centre inputs through a language model that queries healthcare databases and returns responsive information, aimed at pharmacy benefit call handling.

  How it works

  Claim 1 records audio inputs at a call center agent device from members or agents, passes them to LLM circuitry in a neural network, has the LLM develop data requests, accesses several healthcare-related databases for the requested information, identifies the responsive information, and returns it to the agent device to avoid manual searching.

  Where it fits

  Contact-center servicing for a pharmacy benefit manager or health plan.

  Stage

  Filed December 2024, the applicant having responded to an office action, on a new claim presented after the original claims were cancelled, typically following a rejection.

  The claim new claim

  A call center process, comprising:

  1. recording audio inputs at the call center agent device from members, call center agents, or both
  2. providing the audio inputs to a large language model (LLM) circuitry in an artificial neural network
  3. processing the inputs by the LLM circuitry to develop LLM data requests
  4. accessing a plurality of healthcare-related databases using the LLM circuitry and containing information for the LLM data requests
  5. identifying information from the plurality of databases responsive to the LLM data requests
  6. communicating the information back to the call center agent device to prevent the agent from manually searching through different databases in the plurality of databases. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 1/8/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  A further rejection under section 103 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260188504A1 · filed 12/31/2024
• State Farm awaiting examination published 7/2/2026

  Systems and methods for visualization of utility lines

  Proposes locating utility lines by combining LIDAR returns with existing utility records, aimed at establishing where buried or overhead lines run before work near a property begins.

  How it works

  Claim 1 receives LIDAR data from a LIDAR camera and preexisting utility line data, determines a utility line's location from both, builds a 3D model of the surrounding landscape, and displays a representation of that model with a depiction of the utility line.

  Where it fits

  Property risk assessment and inspection support around underground or overhead utilities.

  Stage

  Filed February 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for determination and visualization of a utility line, the method comprising, via one or more processors, sensors, servers, and/or transceivers:

  1. receiving light detection and ranging (LIDAR) data generated from a LIDAR camera
  2. receiving preexisting utility line data
  3. determining a location of a utility line based upon: (i) the received LIDAR data, and (ii) the preexisting utility line data
  4. building a 3-diminsional (3D) model of a landscape surrounding the utility line
  5. displaying a representation of the 3D model of the landscape and a depiction of the utility line. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260187585A1 · filed 2/19/2026
• TIAA final rejection published 7/2/2026

  System, method, and computer-readable medium for assessing and rationalizing technical debt using AI and machine learning analysis with multi-source data integration

  Describes collecting engineering asset data, processing it with learning algorithms, and generating rationalisation actions and reports, aimed at assessing technical debt across a portfolio of systems.

  How it works

  Claim 1 as amended collects data from multiple sources, normalizes it, processes it with an ensemble of AI, generative and reinforcement-learning models to find asset rationalization opportunities, generates proposed actions and a timeline forecast, validates them with adversarial networks, and appends each with multimodal outputs.

  Where it fits

  Internal technology and technical-debt management for the enterprise, no direct insurance function.

  Stage

  Filed December 2024, the applicant having responded to an office action, on a claim already narrowed during prosecution rather than the opening version.

  The claim as amended, showing changes

  A computing system for assessing and rationalizing technical debt, comprising:

  1. a memory having stored thereon computer-executable instructions that, when executed, cause the computing system to: collect data from multiple sources including internal organization information, external technology and business industry insights, previous technical debt assessments, and business capabilities
  2. transform the collected data into a unified intermediate representation, wherein the transformation includes normalization of data formats and encoding of dependency relationships among assets
  3. process the unified intermediate representation using an ensemble of artificial intelligence models, neural networks, generative artificial intelligence models, deep learning algorithms, and reinforcement learning with human feedback to identify potential asset rationalization opportunities
  4. generate, based on the identified asset rationalization opportunities, one or more proposed asset rationalization actions corresponding to software and hardware assets
  5. generate a forecast of timelines for asset rationalization and technical debt reduction based on the one or more proposed asset rationalization actions
  6. validate the one or more proposed asset rationalization actions using adversarial networks, wherein the adversarial networks generate synthetic data
  7. generate one or more multimodal explanatory outputs for the proposed asset rationalization actions, wherein each multimodal explanatory output includes synchronized narrative text, detailed reports, process diagrams, and artificial intelligence-enabled videos
  8. and append each proposed asset rationalization action and its corresponding multimodal explanatory outputs to a versioned repository data structure. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 3 office actions

  First action 3/13/2025, most recent 3/10/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260187571A1 · filed 12/31/2024
• Humana awaiting examination published 7/2/2026

  Machine learning platform and pipeline for efficient data processing

  Describes a shared feature store and standardised training pipeline so models reuse de-sensitised features rather than each team rebuilding them, aimed at reducing duplicated model development work.

  How it works

  Claim 1 generates standardized features from raw external data into a shared feature store, selects input features from it, generates templates for the predictive model, trains models on the selected features and templates, identifies a preferred trained model, registers it and its artifacts, and automates its execution on future data.

  Where it fits

  Model development infrastructure for a health insurer's analytics, not a specific pricing step.

  Stage

  Filed February 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for generating a standardized predictive model using a modeling system, the method comprising:

  1. generating, by a feature generator of the modeling system, a plurality of standardized features from raw data obtained from one or more external data sources, the plurality of standardized features being stored in a centralized feature store accessible to a plurality of machine learning models
  2. selecting a set of input features from the centralized feature store for use by the predictive model
  3. generating, using a template generator of the modeling system, one or more templates for the predictive model
  4. training the predictive model using the selected set of input features and the one or more templates to generate one or more trained models
  5. identifying a preferred trained model from the one or more trained models
  6. registering the preferred trained model and associated model artifacts in a centralized registry
  7. automating execution of the preferred trained model on future data sets. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260187543A1 · filed 2/27/2026
• Equifax awaiting examination published 7/2/2026

  Bias detection and reduction in machine-learning techniques

  Seeks to reduce bias in a risk assessment model by obtaining data on a protected attribute during training and adjusting the model against it, addressing fairness inside the training loop.

  How it works

  Claim 1 trains a risk model on predictor variables and outputs, obtains protected attribute data, computes a bias metric from correlations among a predictor variable, the protected attribute and model outputs, flags bias when it indicates it, modifies and retrains the model, then predicts a risk indicator used to control an entity's access.

  Where it fits

  Fair-lending style bias control in risk scoring, relevant to underwriting model governance.

  Stage

  Filed February 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A method that includes one or more processing devices performing operations comprising:

  1. determining, using a machine learning model trained using a training process, a risk indicator for a target entity from predictor variables associated with the target entity, wherein the risk indicator indicates a level of risk associated with the target entity, wherein the training process includes operations comprising: training the machine learning model using training samples comprising training predictor variables and training outputs corresponding to the training predictor variables, obtaining protected attribute data
  2. calculating a bias metric using the protected attribute data and model-generated data obtained from the machine learning model, wherein the bias metric comprises a correlation metric based at least in part on a first correlation between values of a training predictor variable and the protected attribute data and a second correlation between the model-generated data and the protected attribute data
  3. determining that a bias is detected based on the bias metric
  4. modifying the machine learning model based on the detected bias
  5. re-training the machine learning model
  6. transmitting, to a remote computing device, a responsive message including at least the risk indicator for use in controlling access of the target entity to one or more interactive computing environments. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260187533A1 · filed 2/17/2026
• Equifax awaiting examination published 7/2/2026

  Data protection via attributes-based aggregation

  Proposes obfuscating restricted records by aggregating them on shared attributes, so a requester receives grouped values rather than the underlying sensitive rows.

  How it works

  Claim 1 takes a request to access restricted sensitive records, transforms them with a model chosen by the intended use of the aggregated output, groups the records into aggregation segments using a selected segmentation model, and generates aggregated data by combining the records' individual sensitive attributes across those segments.

  Where it fits

  Privacy-preserving data handling for risk and credit analytics, a data-governance function.

  Stage

  Filed February 2026, docketed and awaiting a first office action, on the original claims as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method, in which one or more processing devices perform operations comprising:

  1. receiving a request to access at least a portion of a plurality of sensitive data records to which access is restricted
  2. transforming the plurality of sensitive data records using a data transformation model, the data transformation model determined based on a targeted use of aggregated data, wherein the aggregated data is generated based on the plurality of sensitive data records
  3. grouping, by executing a data segmentation model selected from a set of available segmentation models, the plurality of sensitive data records into a plurality of aggregation segments
  4. generating the aggregated data by combining individual sensitive attributes of the plurality of sensitive data records based on the plurality of aggregation segments. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260187269A1 · filed 2/18/2026
• Cape Analytics awaiting examination published 6/25/2026

  System and method for property typicality determination

  Describes scoring how typical a property is against a reference population of comparable properties, and identifying which attributes drive that score, aimed at property risk characterisation.

  How it works

  Obtains imagery of a property in a hazard-exposed region, uses a trained model to extract structural and environmental risk attributes into a vector, selects reference properties sharing the same hazard profile, normalizes each attribute by its variance across that reference set, and computes a typicality metric from the normalized vectors.

  Where it fits

  Property underwriting and catastrophe risk assessment, flagging outlier structures against local peers.

  Stage

  Filed February 2026 and awaiting examination; the claim read here was newly substituted for the cancelled original claims, so examination of it has not begun.

  The claim new claim

  A method for generating information for property information, comprising:

  1. obtaining images associated with a property located in an environmental hazard- exposed region
  2. extracting, using a trained machine learning model, attribute values from the images, wherein an attribute value comprises a structural attribute value and an environment risk attribute value associated with the property
  3. determining an attribute vector from the attribute values
  4. determining reference properties for the property based on a shared environment hazard exposure profile
  5. determining reference attribute vectors and corresponding variances of the reference attribute vectors for the reference properties using at least a reference structural attribute value and a reference environment risk attribute value of the reference properties
  6. normalizing the attribute vector and the reference attribute vectors using a respective variance corresponding to a respective attribute value to obtain a normalized attribute vector and normalized reference attribute vectors
  7. determining a typicality metric for the property based on the normalized attribute vector and the normalized reference attribute vectors. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260179361A1 · filed 2/16/2026
• State Farm under examination published 6/25/2026

  Artificial intelligence-based systems and methods for smart home related data predictions and recommendations

  Seeks to extract home data from inspection reports with a model trained on historical correlations, aimed at turning narrative inspection text into structured input for smart home predictions.

  How it works

  Receives a home inspection report, extracts home data with an AI model trained on correlations between past reports and past home data, stores it across defined fields, finds a field with no value, generates a predicted value for it from historical home data, and stores that prediction.

  Where it fits

  Property underwriting and inspection intake, filling gaps in home records used to assess risk.

  Stage

  Filed November 2025 and awaiting examination; the claim read here is the original as filed, the broadest version the applicant will hold.

  The claim as filed

  A computing device utilizing artificial intelligence tools to extract and augment data from a home inspection report and generate home recommendations, the computing device comprising at least one processor and at least one memory device in communication with the at least one processor, the at least one processor configured to:

  1. receive a first home inspection report associated with a first home
  2. extract, using an artificial intelligence model, home data from the first home inspection report, the artificial intelligence model including extraction tools and trained using correlations between historical home inspection reports and historical home data
  3. store the extracted home data for the first home in a data structure including a plurality of data fields
  4. identify at least one data field of the plurality of data fields that is missing a data value
  5. generate, using the artificial intelligence model, at least one predicted data value for the identified at least one data field based upon the historical home data
  6. store the at least one predicted data value in the identified at least one data field. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 9/2/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  A further rejection under section 102 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260179158A1 · filed 11/19/2025
• State Farm awaiting examination published 6/25/2026

  Chatbot for reviewing insurance claims complaints

  Proposes taking insurance claim complaints through a chatbot that categorises each complaint, aimed at triaging complaint handling by type before a human reviewer sees it.

  How it works

  Identifies a website carrying a complaint, scrapes the text, feeds it to a chatbot using a generative pre-trained transformer or an LSTM, categorizes the complaint by either its tone or the policy issue involved, assembles a report noting that category, and sends it to a complaint administrator's device.

  Where it fits

  Claims and complaint servicing, triaging insurance complaints before they reach a human administrator.

  Stage

  Filed February 2026 and awaiting examination; the claim read here is the original as filed and has not yet been tested by an examiner.

  The claim as filed

  A computer-implemented method for using a chatbot (i) implemented by one or more processors, and (ii) including a generative pre-trained transformer (GPT), and/or long-short-term-memory (LSTM) to analyze complaints, comprising:

  1. determining, via the one or more processors, a website with a complaint
  2. scraping, via the one or more processors, the complaint from the website
  3. receiving, using the chatbot, the complaint, wherein the chatbot includes the GPT, and/or the LSTM
  4. categorizing, using the chatbot, the complaint by determining a category of the complaint, the category comprising a tone category or a policy category
  5. building, using the chatbot, a complaint report including information of the complaint and an indication of the category
  6. sending, using the chatbot, the complaint report to a complaint administrator computing device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260179152A1 · filed 2/12/2026
• State Farm awaiting examination published 6/25/2026

  Customized data management systems and methods with artificial intelligence platform

  Describes monitoring a linked user account and acting on the data it returns, with an artificial intelligence platform driving what the monitoring tool does on the account holder's behalf.

  How it works

  Opens a link to a user's device, ingests account data, runs a management tool whose user-specific AI model is trained on that user's historical data, detects an account event that needs handling, generates a response satisfying the event's condition without any contemporaneous input from the user, and transmits it back.

  Where it fits

  Policy and account servicing, automating responses to account events for insurance customers.

  Stage

  Filed December 2025 and awaiting examination; the claim read here is the original as filed, the broadest version on record.

  The claim as filed

  1. A computer system for providing customized data management (CDM), the computer system comprising at least one processor and at least one memory device in communication therewith, the at least one processor in further communication with one or more user computer devices, the at least one processor programmed to: establish a communication link with a computing device associated with at least one user account of a plurality of user accounts associated with a user
  2. receive user data from the at least one user account via the communication link
  3. execute a CDM tool to actively monitor the at least one user account, wherein the CDM tool includes a user-specific artificial intelligence (AI) model trained based upon historical user data associated with the user and the at least one user account
  4. detect, by the CDM tool and the received user data, an event associated with the at least one user account and that the least one user account requires a managed response to the event
  5. in response to the detected event, generate the managed response using the user-specific AI model and without contemporaneous input from the user, the managed response satisfying a condition associated with the detected event
  6. transmit, via the communication link, the managed response to the computing device associated with the at least one user account. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260179100A1 · filed 12/4/2025
• Humana awaiting examination published 6/25/2026

  Machine learning based model for determining effective communication mechanism with users

  Describes building a feature vector from a user profile that includes past communications and responses, then selecting which channel to use for the next contact.

  How it works

  Builds a feature vector from a member's profile of past communications and actions, runs a separate model per channel to score how likely the member acts on a message there, each model trained by labeling users responsive or not from their communication and event histories, then selects the top-scoring channel and sends via that channel's API.

  Where it fits

  Health plan member outreach, choosing the contact channel most likely to prompt an action such as care follow-up.

  Stage

  Filed February 2026 and awaiting examination; the claim read here is the original as filed and remains untested by an examiner.

  The claim as filed

  A computer-implemented method for selecting a communication channel for communicating with users, comprising:

  1. identifying, by one or more processors, a user for sending a communication via one of a plurality of communication channels
  2. accessing, by the one or more processors, a user profile of the user stored in a user data store, the user profile comprising time series data based on past communications with the user and past user actions
  3. generating, by the one or more processors, a feature vector based on the user profile of the user
  4. for each communication channel of the plurality of communication channels: providing the generated feature vector as input to a machine learning based model associated with the communication channel, the machine learning based model configured to receive a feature vector describing an input user and predict a likelihood of the input user performing an expected user action responsive to a communication sent via the communication channel
  5. executing the machine learning based model associated with the communication channel to determine a score indicating a likelihood of the user performing the expected user action responsive to receiving a communication via the communication channel
  6. wherein the machine learning based model associated with the communication channel is trained using training data generated by: selecting a set of users
  7. for each user from the set: accessing a communication time series representing communications performed to the user at various time points using the communication channel
  8. accessing an event time series representing events indicating instances of expected user actions performed by the user at various time points
  9. determining whether the user is responsive to communications performed using the communication channel based on the communication time series and the event time series
  10. labeling the user based on the determination
  11. selecting, by the one or more processors, a communication channel from the plurality of communication channels based on the scores determined by the machine learning based models
  12. sending a communication to the user via the selected communication channel by invoking a channel-specific application programming interface of a communication engine associated with the selected communication channel. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260178941A1 · filed 2/17/2026
• Clover Health awaiting examination published 6/18/2026

  Clinical assessment tool

  Describes identifying providers and locations able to perform a referred procedure, aimed at steering a referral toward capable sites at the point a provider submits it.

  How it works

  Receives a referral for a patient to undergo a procedure, determines candidate medical locations or providers weighted by patient location, insurance network, or cost, surfaces those options on the referring provider's display, takes the user's selection, and processes the referral against the selected location or provider.

  Where it fits

  Health plan care management and provider steerage, guiding referrals toward in-network, lower-cost options.

  Stage

  Filed November 2025 and awaiting examination; the claim read here is the original as filed, the broadest version the applicant will hold.

  The claim as filed

  A method comprising:

  1. receiving, via a computing device associated with a first medical provider, a first indication of a referral for a patient to undergo a procedure
  2. determining at least one of one or more medical locations or one or more medical providers associated with the procedure, the at least one of one or more medical locations or one or more medical providers being based at least in part on at least one of a patient location, an insurance network, or a cost associated with the procedure
  3. cause at least one title associated with the at least one of the one or more medical locations or the one or more medical providers to surface on a display of the computing device
  4. receiving, from the computing device, a second indication of selection of at least one of a medical location of the medical locations or a medical provider of the one or more medical providers
  5. processing the referral for the patient to undergo the procedure based at least in part on the at least one of the medical location or the medical provider. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260171256A1 · filed 11/17/2025
• Clover Health awaiting examination published 6/18/2026

  Machine learning models for gaps in care and medication actions

  Describes models that flag medication issues and gaps in a member's care, then propose actions to close them, aimed at intervening before a gap becomes a claim.

  How it works

  Trains machine learning models to flag gaps in care or medication actions, pulls a patient's data from multiple medical databases, formats it into model features, and outputs each finding with a confidence value, then on learning the patient will be seen within a period notifies the provider of the gap or action before or during that visit.

  Where it fits

  Health plan quality and care-gap closure, prompting providers at the point of an upcoming encounter.

  Stage

  Filed November 2025 and awaiting examination; the claim read here is the original as filed and has not been examined.

  The claim as filed

  A system comprising:

  1. one or more processors
  2. non-transitory computer-readable media storing first computer-executable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising: generating machine learning models configured determine at least one of gaps in medical care or one or more actions associated with medication
  3. receiving first data associated with a medical patient from multiple sources via a computing network, the multiple sources each associated with medical-related database
  4. formatting the first data into model features configured to be input into the machine learning models, wherein individual ones of the machine learning models are trained to receive the model features and output second data indicating a probability that the at least one of the gaps in medical care are identified or the one or more actions should occur
  5. inputting the model features into the machine learning models
  6. generating, utilizing at least the machine learning models, the second data and a confidence value associated with the second data
  7. receiving, from a computing device executing an application configured to display a graphical user interface associated with the gaps in medical care and the one or more actions, an indication that the medical patient will be seen by a medical service provider during a period of time
  8. sending, to the computing device and at least one of before or during the period of time, a notification including at least one of identification of the gap in medical care or the one or more actions as determined from the machine learning models. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260171247A1 · filed 11/14/2025
• State Farm awaiting examination published 6/18/2026

  Interactive video accessibility compliance systems and methods

  Proposes scanning a video file for interactive elements and prompting for the inputs needed to describe them, aimed at bringing interactive video into accessibility compliance.

  How it works

  Analyzes a video executable file to locate where interactive elements will occur during playback, receives label data tying a label to each such location, links every label and its interactive element to the corresponding point in the video, and outputs an enhanced video file with those labels embedded.

  Where it fits

  General accessibility tooling for interactive video; the claim does not tie the mechanism to a specific insurance workflow.

  Stage

  Filed February 2026 and awaiting examination; the claim read here is the original as filed, the broadest version on record.

  The claim as filed

  A video accessibility computing system for generating interactive videos, the video accessibility computing system comprising a processor in communication with at least one memory, the processor configured to:

  1. analyze a first video executable file associated with an interactive video to identify one or more locations within the first video executable file where one or more interactive elements will occur during playback of the interactive video
  2. receive label data for one or more labels each corresponding to the one or more locations within the interactive video, each label of the one or more labels being associated with a corresponding interactive element of the one or more interactive elements
  3. process the label data to link each label of the one or more labels and the corresponding interactive element of the one or more interactive elements to a respective location of the one or more locations within the interactive video
  4. generate and output an enhanced video executable file including the linked labels embedded therein which are linked to the associated one or more interactive elements, wherein the enhanced video executable file is configured to be executed in association with the playback of the interactive video. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260171121A1 · filed 2/6/2026
• TIAA final rejection published 6/18/2026

  Automated technical debt evaluation and scoring using artificial intelligence

  Describes deriving coding signals from engineering data and scoring them against business capabilities to produce a technical debt score for a portfolio of systems.

  How it works

  Trains a neural network on past coding practices and outcomes, normalizes data collected across a technology ecosystem, derives coding signals from code-quality metrics and business capabilities, assigns them weighted scores, classifies those into ranges, then prompts a fine-tuned language model to produce a report explaining the debt score.

  Where it fits

  Enterprise software risk management; the claim describes internal IT governance rather than an insurance workflow.

  Stage

  Filed December 2024 and now under a final rejection; the claim read here was amended during prosecution, meaning the applicant narrowed it, not that the examiner allowed it.

  The claim as amended, showing changes

  A computing system for evaluating and managing technical debt within a technology ecosystem, the system comprising:

  1. one or more processors
  2. one or more memories storing instructions that, when executed by the one or more processors, cause the system to: train, via the one or more processors, a neural network model using historical data, wherein the historical data includes coding practices and outcomes from previous technical debt assessments
  3. collect, via the one or more processors, data from a technology ecosystem, the data including software and hardware platforms, code repositories, architecture diagrams, tools, and processes
  4. clean, normalize, and categorize the collected data from the technology ecosystem to identify and retain only data points relevant to technical debt assessment
  5. determine, via the one or more processors, coding signals from the collected data, wherein the codingsignals are based on code quality metrics and business capabilities
  6. receive, via the one or more processors, inputs from one or both of (i) a technical debt assessment system and (ii) an organizational redesign system, to inform a technology ecosystem assessment
  7. assign, via the one or more processors, weighted scores to the coding signals based on their impact on technical debt accumulation, wherein the weighted scores are determined based at least in part on expert knowledge and industry best practices
  8. classify, via the trained neural network model, the weighted coding signals and their weighted scores into score ranges to produce classified coding signals, wherein the scoring ranges are indicative of codequality, complexity, and adherence to best practices
  9. determine, via the trained neural network model, non-linear relationships between the classified codingsignals and the business capabilities
  10. apply, via the one or more processors, a neural network to the collected data to identify non-linear relationships between the coding signals and the business capabilities
  11. provide structured instructions and prompts to a natural lan guag e generation model fine- tuned for the technical debt assessment, the structured instructions and prompts that facilitate the natural lan guag e generation model in generatingrelevant analyses and responses
  12. generate, via the natural lan guag e generation model and based on the classified codingsignals, a natural lan guag e report, wherein the natural lan guag e report includes reasoningfor the assigned score ranges, identification of codingpatterns contributing to technical debt accumulation, and recomme

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 4 office actions

  First action 4/28/2025, most recent 7/21/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  References cited against it: Beachus et al, Nerurkar et al, Kreamer et al, Shani et al.

  A further rejection under section 101 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260170555A1 · filed 12/13/2024
• State Farm awaiting examination published 6/18/2026

  Systems and methods for an artificial intelligence-based appliance end-of-life calculator

  Seeks to predict an appliance's remaining lifetime from data about that appliance, aimed at anticipating failure in a home before it becomes a claim.

  How it works

  Receives historical data on many appliances including their observed lifetimes, some of it generated by sensors monitoring appliance parameters, trains an AI model to output a predicted remaining lifetime from an appliance's data, and stores the trained model for later use.

  Where it fits

  Homeowners loss prevention and servicing, anticipating appliance failures that drive property claims.

  Stage

  Filed February 2026 and awaiting examination; the claim read here is the original as filed and covers only training and storing the model.

  The claim as filed

  A computing device for training an artificial intelligence (AI) model to predict a lifetime of one or more appliances, the computing device comprising at least one processor and at least one memory device in communication with the at least one processor, the at least one processor configured to:

  1. receive historical appliance data including historical lifetimes of a plurality of appliances, at least some of the historical appliance data generated by sensors configured to monitor one or more parameters of at least one of the plurality of appliances
  2. train an AI model using the historical appliance data including the historical lifetimes of the plurality of appliances, the AI model configured to output a predicted remaining lifetime of an appliance based upon an input of appliance data relating to the appliance
  3. store the trained AI model in the at least one memory device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260169880A1 · filed 2/5/2026
• Mitchell final rejection published 6/11/2026

  Method for medical record data extraction and summarization using large language artificial intelligence models

  Seeks to extract and summarise unstructured medical records with a large language model, aimed at turning clinical documents into reviewable structured content for claim handling.

  How it works

  Receives a set of medical record documents, pre-processes them into structured medical information, then queries a large language model for a summary using a sequence of successively more specific queries, each shaped by the summary the model returned on the previous pass, and hands the result to downstream workflows.

  Where it fits

  Casualty and medical claims handling, condensing records for adjusters and bill review.

  Stage

  Filed December 2024 with a response now on file; the claim read here was amended during prosecution, so the applicant has narrowed it rather than won allowance.

  The claim as amended, showing changes

  A method for extracting and summarizing unstructured data from medical records, the method comprising:

  1. receiving, by a processor of a medical records processing system, a set of documents comprising medical records
  2. pre-processing, by a processor of the medical records processing system, the received set of documents into a set of medical information
  3. querying, by a processor of the medical records processing system, a Large Language Model (LLM) for summary information of the set of relevant medical information, wherein querying the LLM for summary information of the set of medical information comprises iteratively providing, to the LLM, a sequence of successively more specificqueries related to the set of medical information and based on summary information returned by the LLM in a previous iteration
  4. and providing, by a processor of the medical records processing system, the summary information to one or more workflows for further processing. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 1/15/2026, most recent 9/3/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Madan, Brigham.

  A further rejection under section 101 and 102 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260162783A1 · filed 12/6/2024
• State Farm awaiting examination published 6/11/2026

  Artificial Intelligence (AI) for Prediction and/or Prevention of Home Loss and/or Damage

  Describes constructing a customised training dataset before training a model, aimed at predicting and preventing home loss where general-purpose training data fits the risk poorly.

  How it works

  Builds a customer-specific training set by dropping records from a base insurance dataset based on other customers' geographic distance from the target customer and on a temporal constraint, trains a model on that set, and runs the customer's data through it to produce a determination such as probability of loss or an indemnity estimate.

  Where it fits

  Property underwriting and loss prevention, localizing risk models to a customer's area.

  Stage

  Filed December 2025 and awaiting examination; the claim read here is the original as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for training and using a machine learning (ML) model to make an insurance-related determination, the computer-implemented method comprising:

  1. constructing, via one or more processors, a customized training dataset by: receiving a geographic location of a customer
  2. receiving a base insurance dataset including data of a plurality of insurance customers
  3. building the customized training dataset by removing data from the base insurance dataset based upon: (i) respective geographic distances between the geographic location of the customer and respective insurance customers of the plurality of insurance customers, and (ii) a temporal constraint
  4. training, via the one or more processors, the ML model by inputting the customized training dataset into the ML model
  5. determining, via the one or more processors, by inputting data of the customer into the trained ML model, one or more of: (i) probability of a loss by cause of loss, (ii) a cost estimate by cause of loss, (iii) probability of loss by loss-comment-code, (iv) indemnity estimate by loss-comment-code, (v) percent change in probability of loss given performed insight, (vi) probability that customer will perform insight, (vii) estimated cost of performed insight, (viii) customer segmentation, and/or (ix) probability of the customer placing an insurance claim. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260162188A1 · filed 12/11/2025
• State Farm awaiting examination published 6/11/2026

  Method of controlling for undesired factors in machine learning models

  Describes training a model to read an applicant image but exclude factors such as age, sex, ethnicity or race, then set a life or health premium from only the remaining features, aimed at keeping protected characteristics out of an underwriting decision.

  How it works

  Trains a neural network on images to correlate appearance with personal or health characteristics, identifies undesired factors such as age, sex, ethnicity, or race along with their interaction terms, then analyzes an applicant's image while excluding those factors and suggests an insurance premium based only on the remaining characteristics.

  Where it fits

  Life and health underwriting, pricing from an applicant image while attempting to control for protected characteristics.

  Stage

  Filed April 2025 and awaiting examination; the claim read here is the original as filed and has not been examined.

  The claim as filed

  A computer-implemented method for training and using a neural network to evaluate an insurance applicant as part of an underwriting process to determine an appropriate insurance premium, wherein the neural network controls for consideration of one or more undesired factors which might otherwise be considered by the neural network, the computer-implemented method comprising, via one or more processors :

  1. training the neural network to probabilistically correlate an aspect of appearance with a personal and/or health-related characteristic by providing the neural network with a training data set of images of individuals having known personal and/or health-related characteristics, including the one or more undesired factors
  2. identifying the one or more undesired factors
  3. identifying one or more relevant interaction terms between the one or more undesired factors
  4. receiving via a communication element an image of the insurance applicant
  5. analyzing with the neural network the image of the insurance applicant to probabilistically determine the personal and/or health-related characteristics for the insurance applicant, wherein such analysis excludes the identified one or more undesired factors
  6. suggesting with the neural network the appropriate insurance premium based at least in part on the probabilistically determined personal and/or health-related characteristics but not on the one or more undesired factors to control for undesired prejudice or discrimination in neural networks. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260162184A1 · filed 4/18/2025
• LexisNexis Risk awaiting examination published 6/4/2026

  Computer vision learning and object detection for long-tailed data distributions

  Describes detecting objects in long-tailed distributions using pseudo-labelled unlabelled data, aimed at recognising rare categories without assembling a large labelled set for each one.

  How it works

  Splits a dataset into frequent head classes and rare tail classes, pre-trains a detection model on the head classes, adapts it to the tail by pseudo-labeling unlabeled images and initializing tail parameters from the pre-trained model, then fine-tunes on a balanced dataset while updating the classifier and regressor and freezing the backbone.

  Where it fits

  General computer-vision tooling; the claim does not tie the method to a specific insurance workflow, though the filer serves risk analytics.

  Stage

  Filed April 2025 and awaiting examination; the claim read here is the original as filed, the broadest version on record.

  The claim as filed

  A computer-implemented method for long-tailed object detection, comprising:

  1. obtaining a dataset with labeled head classes appearing in more than a threshold number of samples and labeled tail classes appearing in fewer than the threshold number of samples
  2. pre-training a detection model on head class images of the dataset, wherein the detection model comprises: a backbone network for feature extraction
  3. a detection head comprising a classifier module and a regressor module for bounding box predictions
  4. augmenting the pre-trained detection model by pseudo-labeling unlabeled images and adapting the detection model to tail classes by initializing tail parameters of the detection model from the pre-trained detection model and fine-tuning with tail class images
  5. fine-tuning the detection model on a balanced dataset comprising both head and tail classes, wherein the fine-tuning updates the classifier module and the regressor module while freezing parameters of the backbone network. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260154953A1 · filed 4/7/2025
• State Farm awaiting examination published 6/4/2026

  Systems and methods for analyzing and mitigating community-associated risks

  Seeks to combine sensor readings with building records in a trained model to characterise risks around a structure, aimed at identifying community-level exposure rather than a single property.

  How it works

  Trains a model on the layout of area security systems, historical scheduled events, and past security occurrences, takes the location and time of a new scheduled event, predicts a potential security occurrence, identifies a person whose commute passes near that location, and sends them a risk-mitigating recommendation timed to the event.

  Where it fits

  Personal-lines risk mitigation and loss prevention, warning insureds away from predicted incident locations.

  Stage

  Filed January 2026 and awaiting examination; the claim read here is the original as filed and remains untested by an examiner.

  The claim as filed

  1. A computer system for analyzing data associated with scheduled events at geographic locations using a machine learning model and outputting risk mitigating recommendations, the computer system comprising at least one processor and at least one memory device, the at least one processor programmed to: train a machine learning model using at least i) historical systems data comprising a layout of security systems in a geographic area, ii) historical scheduled event data comprising historical scheduled events in the geographic area, and iii) one or more historical security occurrences associated with the historical scheduled events
  2. receive new systems data from at least one services computer system, the new systems data identifying a geographic location of a scheduled event in the geographic area and a time that the scheduled event is scheduled to occur
  3. input the new systems data into the machine learning model
  4. receive an output from the machine learning model that identifies at least one potential security occurrence associated with the scheduled event
  5. identify at least one person potentially impacted by the at least one potential security occurrence based at least in part upon the geographic location of the scheduled event and at least one individual profile associated with the at least one person indicating that at least one scheduled or predicted commute of the at least one person includes travel proximate to the geographic location
  6. transmit security data associated with the at least one potential security occurrence and including a risk mitigating recommendation to a computing device associated with the at least one person, wherein the security data causes initiation of at least one recommended action to be taken proximate to the time that the scheduled event is scheduled to occur, and wherein the at least one recommended action increases a likelihood of the person avoiding the at least one potential security occurrence. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260154769A1 · filed 1/29/2026
• Equifax awaiting examination published 6/4/2026

  Power graph convolutional network for explainable machine learning

  Describes a graph convolutional network that produces a risk indicator alongside an account of which predictors drove it, aimed at risk scoring that can be explained to a regulator.

  How it works

  Runs an entity's predictor variables through a power graph convolutional network whose convolutional layer applies an adjacency weight matrix and whose dense layer applies a weight vector, trains those weights under an explainability constraint, produces a risk indicator, and transmits it to control the entity's access.

  Where it fits

  Credit and identity risk scoring, where an explainable score gates access decisions.

  Stage

  Filed April 2024 and awaiting examination; the claim read here is the original as filed, the broadest version the applicant will hold.

  The claim as filed

  A method that includes one or more processing devices performing operations comprising:

  1. determining, using a power graph convolutional network trained by a training process, a risk indicator for a target entity from predictor variables associated with the target entity, wherein: the power graph convolutional network comprises (a) a convolutional layer configured to apply an adjacency weight matrix on the predictor variables to generate modified predictor variables and (b) a dense layer configured to apply a weight vector on the modified predictor variables to generate the risk indicator, and the training process comprises adjusting a first set of weights in the adjacency weight matrix and a second set of weights in the weight vector based on a loss function of the power graph convolutional network and under an explainability constraint on the first set of weights or the second set of weights of the power graph convolutional network
  2. transmitting, to a remote computing device, a responsive message including at least the risk indicator for use in controlling access of the target entity to one or more interactive computing environments. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260154532A1 · filed 4/4/2024
• Allstate awaiting examination published 6/4/2026

  Systems and methods for user classification with respect to a chatbot

  Describes a chatbot that classifies a user into one of two levels from attributes of an open-ended query, then handles higher-level users through further open-ended follow-ups and lower-level users through closed-ended prompts to gather intents before completing a task.

  How it works

  Takes an open-ended query and classifies the user as a first or second level depending on whether attributes of the query cross a threshold; a first-level user is asked further open-ended questions before a fulfillment task runs, while a second-level user gets closed-ended questions to collect the missing intents or entities.

  Where it fits

  Customer-facing chatbot triage; the abstract describes conversational routing, not a specific insurance transaction.

  Stage

  Filed July 2025, published June 2026 and awaiting examination; this claim was newly substituted after the original claims were cancelled, not a granted right.

  The claim new claim

  A system comprising:

  1. a computing device comprising a processor and a non-transitory computer readable memory, the computing device configured to: receive, from an input device, an open-ended query from a user
  2. classify the user as either a first user level or a second user level based on one or more attributes of the open-ended query, wherein the user is classified as the first user level based on the one or more attributes exceeding a threshold associated with a respective one of the one or more attributes and the user is classified as the second user level based on the one or more attributes not exceeding the threshold associated with a respective one of the one or more attributes
  3. in response to classifying the user as the first user level: prompt the user with one or more additional open-ended inquiries to obtain one or more additional intents or one or more additional entities, and execute a fulfillment task based on the one or more additional intents or the one or more additional entities
  4. in response to classifying the user as the second user level, prompt the user with one or more closed-ended inquiries to obtain one or more additional intents or one or more additional entities from the user. New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260154513A1 · filed 7/11/2025
• Genpact under examination published 6/4/2026

  Multi-modal image extraction and retrieval using retrieval augmented generation

  Describes a retrieval-augmented generation system that replaces an image in a document with a localization tag marking where it sat, then vectorizes and stores the tagged document so the image can later be searched and retrieved.

  How it works

  A tagging engine locates an image within an electronic document, generates a localization tag recording its position, and swaps the image for that tag; a vector engine then converts the modified document into vectors and stores it in a vector database so the content can be searched and retrieved through retrieval-augmented generation.

  Where it fits

  General-purpose document and image retrieval; the abstract names no insurance workflow.

  Stage

  Filed December 2024, published June 2026 and awaiting examination; the claim read here is the original as filed, the broadest version an examiner is most likely to reject.

  The claim as filed

  A system for retrieval of images using Retrieval Augmented Generation (RAG), comprising:

  1. a tagging engine configured to: receive an electronic document having an image
  2. determine a location of the image in the electronic document
  3. generate an image localization tag (ILT) based on the location of the image
  4. replace the image in the electronic document with the ILT to produce a modified electronic document
  5. a vector engine configured to vectorize and store the modified electronic document in a vector database for subsequent search and retrieval using RAG. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 9/4/2026. The grounds raised so far:

  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Dar et al, Lyons et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260154332A1 · filed 12/3/2024
• Mitchell awaiting examination published 5/28/2026

  Methods for managing one or more uncorrelated elements in data and devices thereof

  Describes selecting a diagnostic mapping table from the code and data format on an electronic claim, aimed at reconciling codes that do not correspond across claim systems.

  How it works

  Receives an electronic claim with a diagnostic code in one industry's data format, selects the matching mapping table, resolves the code to a body part and its laterality, feeds those with a categorization table from another format into a trained model to produce an assessment rating, and initiates an action on the claim.

  Where it fits

  Casualty and medical claims adjudication, reconciling diagnostic coding across different industry systems.

  Stage

  Filed January 2026 and awaiting examination; the claim read here is the original as filed and has not been examined.

  The claim as filed

  A method comprising:

  1. receiving, by a processor of an automated claims processing system, from a client device, a request to process an electronic claim comprising a diagnostic code associated with a treatment procedure in one of a plurality of data environment formats, wherein the plurality of data environment formats comprises at least two data environment formats in different industry types
  2. retrieving, by the processor of the automated claims processing system, the electronic claim specified by the request
  3. identifying, by the processor of the automated claims processing system, one of a plurality of diagnostic mapping tables based on the diagnostic code in the one of the plurality of data environment formats in the electronic claim, wherein each data environment format of the plurality of data environment formats is associated with one of the plurality of diagnostic mapping tables
  4. determining, by the processor of the automated claims processing system, first and second identifiers corresponding to the diagnostic code in the one of the plurality of data environment formats based on the identified one of diagnostic mapping tables, wherein the first identifier represents at least one of a plurality of human body parts and the second identifier represents laterality of the at least one human body part
  5. determining, by the processor of the automated claims processing system, one of a plurality of assessment ratings based on input parameters, wherein the input parameters comprise at least one of the diagnostic code, the determined one of the plurality of the human body parts specified by the first identifier, the determined laterality specified by the second identifier, and a selected categorization table associated with another one of the plurality of data environment formats, wherein the selected categorization table comprises one of a plurality of categorization tables, each categorization table associated with one of the plurality of data environments, wherein the selected categorization table is associated with a data environment other than the data environment of the diagnostic code, wherein determining the one of a plurality of assessment ratings based on input parameters comprises applying the input parameters as inputs to a trained machine learning model, wherein responsive to the inputs, the trained machine learning model outputs the one of a plural

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260148820A1 · filed 1/22/2026
• State Farm awaiting examination published 5/28/2026

  Artificial Intelligence for Sump Pump Monitoring and Service Provider Notification

  Proposes detecting a failing sump pump and having a chatbot solicit repair or replacement quotes, aimed at closing the loop between a sensor fault and a service booking.

  How it works

  Takes operating data from sump pump sensors that detect water or power interruption, polls a weather server for conditions at the pump's location, merges the two into a dataset, sends it with a prompt to a machine learning chatbot that generates a natural-language fault-detection analysis, and presents that analysis.

  Where it fits

  Homeowners loss prevention, catching sump pump failures that lead to water-damage claims.

  Stage

  Filed January 2026 and awaiting examination; the claim read here is the original as filed, the broadest version on record.

  The claim as filed

  A computer system for sump pump monitoring and repair service provider notification, the computer system comprising:

  1. one or more processors
  2. a memory storing executable instructions thereon that, when executed by the one or more processors, cause the one or more processors to: receive location and operating data indicative of an operating status of the sump pump from one or more sump pump sensors, wherein the one or more sump pump sensors are configured to detect water or power interruption, wherein the operating status corresponds to a water detection or power interruption event
  3. automatically poll a weather server to obtain weather data pertaining to the location of the sump pump
  4. integrate the operating data and weather data to generate a dataset
  5. transmit the dataset and a prompt for sump pump fault detection to a machine learning (ML) chatbot to cause an ML model to generate a natural language sump pump fault detection analysis from the dataset
  6. present the natural language fault detection analysis. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260148320A1 · filed 1/15/2026
• State Farm awaiting examination published 5/28/2026

  Intelligent user interface monitoring and alert

  Describes monitoring how a user works through an interface and recommending actions to make those interactions more efficient.

  How it works

  Receives an indication of a fault tied to a user interface, reviews the recorded sequence of user inputs, identifies the first input that produced the fault, determines a different input that would have prevented it, and sends the device a notification recommending that alternative input.

  Where it fits

  General user-interface support; the claim does not tie the mechanism to a specific insurance workflow.

  Stage

  Filed January 2026 and awaiting examination; the claim read here was amended during prosecution, meaning the applicant narrowed it, not that an examiner has agreed.

  The claim as amended, showing changes

  A computing system comprising:

  1. a computer-readable media storing instructions which, when executed by the processor, cause the processor to: receive, from a computing device, an indication of a fault associated with a user interface
  2. access event data associated with the user interface, the event data indicative of a first user inputindicating a sequence of user inputs provided to the user interface
  3. determine, based on the event data and the fault, a first user input of the sequence of user inputs that resulted in the fault
  4. determine, based on the fault and the first user input, a second user input, different from the first user input, that preventswould prevent the fault
  5. provide, to the computing device, a notification indicating the second user input. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260148311A1 · filed 1/21/2026
• State Farm under examination published 5/28/2026

  Systems and methods for an artificial intelligence-based appliance end-of-life calculator

  Describes an AI model trained on historical lifetimes of similar appliances that would predict a selected appliance's remaining life from retrieved appliance data and present a repair-or-replace recommendation to the user.

  How it works

  Presents a user interface to select an appliance, retrieves that appliance's data, computes a predicted remaining lifetime with an AI model trained on historical lifetimes of similar appliances, generates a recommendation to repair or replace it, and displays the lifetime and recommendation.

  Where it fits

  Homeowners servicing and loss prevention, guiding appliance replacement before a failure causes a claim.

  Stage

  Filed September 2025 and under examination; the claim read here is the original as filed, the broadest version the applicant will hold.

  The claim as filed

  A computing device for monitoring and predicting a lifetime of one or more appliances, the computing device comprising at least one processor and at least one memory device, the at least one processor configured to:

  1. cause a user device to present a user interface prompting a selection of an appliance
  2. receive, from the user device, a selection of a first appliance
  3. retrieve appliance data relating to the first appliance
  4. compute, using an artificial intelligence model, a predicted remaining lifetime of the first appliance based upon the retrieved appliance data of the first appliance, wherein the artificial intelligence model is trained based upon historical appliance data including data associated with historical lifetimes of similar appliances
  5. generate a recommendation to repair or replace the first appliance based upon the predicted remaining lifetime
  6. cause the user interface to present at least the predicted remaining lifetime of the first appliance and the generated recommendation. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/1/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  References cited against it: Gibson et al, US 10,013,677, Gibson, US 12,394,031.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260148199A1 · filed 9/26/2025
• State Farm awaiting examination published 5/28/2026

  Machine learning systems and methods for generating calendar event data from one or more input data types

  Describes passing mixed input data through a model that identifies event details within it and turns them into calendar entries.

  How it works

  Feeds mixed-type input data to a machine learning model that identifies and extracts the details defining an event, detects that those details only partially define it, determines what additional parameters are needed, retrieves them by searching further data sources, and generates a calendar event from the combined parameters.

  Where it fits

  General scheduling and productivity tooling; the claim does not tie the mechanism to a specific insurance workflow.

  Stage

  Filed October 2025 and awaiting examination; the claim read here is the original as filed and remains untested by an examiner.

  The claim as filed

  1. A computer system for generating calendar events using machine learning tools and input data that includes one or more data types received over one or more channels of communication, the computer system comprising at least one processor in communication with at least one memory and in further communication with one or more user computing devices, the at least one processor programmed to : receive input data from a user device of the one or more user computing devices, wherein the input data includes one or more data types
  2. input the input data into a machine learning model to identify at least some event data included within the input data, the event data associated with defining an event
  3. cause the machine learning model to extract the event data from the input data
  4. using the machine learning model, determine that the event data includes event parameters that only partially define the event
  5. using the machine learning model, identify additional event data that includes additional event parameters needed to further define the event
  6. cause the machine learning model to retrieve at least one of the additional event parameters by searching additional data sources for the additional event parameters
  7. generate a calendar event based upon the event parameters and the additional event parameters . 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260148197A1 · filed 10/15/2025
• State Farm awaiting examination published 5/21/2026

  Systems and methods for enhanced virtual reality interactions

  Describes presenting a user's item in a shared virtual environment and generating a simulated event involving that item from sensor data, so two users' devices experience the same simulation. The abstract states no insurance function.

  How it works

  Presents a virtual environment containing an item belonging to a first user, receives sensor data about that item, selects a simulated event involving it based on user input, generates the event using the sensor data, and displays it in the virtual environment to both the first user's and a second user's devices.

  Where it fits

  General virtual-environment simulation; the claim does not tie the mechanism to a specific insurance workflow.

  Stage

  Filed August 2025 and awaiting examination; the claim read here is the original as filed, the broadest version on record.

  The claim as filed

  1. A computer system for conducting interactions of a simulated event for a plurality of users in a virtual environment, the computer system comprising at least one processor and at least one memory device in communication with the at least one processor and one or more user computer devices, the at least one processor programmed to: communicate with the one or more user computer devices to cause the one or more user computer devices to present the virtual environment including at least one item of a first user associated with a first user device of the one or more user computer devices
  2. receive sensor data from one or more sensors associated with the first user device, the sensor data including data related to the at least one item
  3. select a simulated event involving the at least one item based upon an input from the one or more user computer devices
  4. generate the simulated event involving the at least one item and using the received sensor data for display within the virtual environment
  5. cause the simulated event involving the at least one item to be displayed within the virtual environment to the first user using the first user device and at least a second user associated with a second user device of the one or more user computer devices, wherein the first user and the second user visually experience the simulated event involving the at least item via the respective first user device and the second user device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260141644A1 · filed 8/27/2025
• CSAA awaiting examination published 5/21/2026

  Systems and Methods for 3D Accident Reconstruction

  Seeks to build a three-dimensional accident reconstruction from photographs of a vehicle, aimed at reconstructing what happened from claim images rather than from an on-site inspection.

  How it works

  Receives photos of a vehicle from a mobile device, generates an accident-reconstruction model with machine learning models that combine the images with the vehicle's technical specifications, has the model indicate the extent of damage for each region, identifies potential damage, and sends instructions to display it graphically.

  Where it fits

  Auto physical-damage claims, estimating and visualizing vehicle damage from photos.

  Stage

  Filed November 2025 and awaiting examination; the claim read here was newly presented in place of the cancelled original claims, so examination of it has not begun.

  The claim new claim

  An accident scene reconstruction system configured for use with a vehicle, the accident scene reconstruction system comprising:

  1. a mobile computing device, wherein the mobile computing device comprises a camera, a network interface, and a graphical user interface
  2. a modeling computing device, wherein the modeling computing device comprises a processor and a non-transitory computer-readable medium, having stored thereon program instructions that, upon execution by the processor, cause the modeling computing device to perform a set of operations comprising: receiving a plurality of images of a particular vehicle from the mobile computing device
  3. generating an accident reconstruction model using one or more machine learning models, wherein the one or more machine learning models are configured to generate the accident reconstruction model using the plurality of images and technical specification materials associated with the particular vehicle, and wherein the accident reconstruction model indicates, for each of multiple regions on the particular vehicle, a respective extent of damage to the particular vehicle
  4. identifying potential damage to the particular vehicle based on at least the accident reconstruction model
  5. transmitting, to the mobile computing device, instructions that cause the mobile computing device to display, via the graphical user interface of the mobile computing device, a graphical indication of the potential damage to the particular vehicle. New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260141632A1 · filed 11/19/2025
• Sedgwick in examination published 5/21/2026

  Computer method and system for applying generative artificial intelligence to insurance data

  Seeks to apply generative techniques to captured insurance claim data and produce content used in handling the claim inside a claims application.

  How it works

  Captures claim data from an insurer, generates prompts specifying which AI models to use with the claim data, digitizes the data for a generative AI system, invokes that system through an API call to produce content responsive to the prompts, and stores and returns the generated content to the claims application for further processing.

  Where it fits

  Insurance claims processing, injecting generative AI output into the claim handling workflow.

  Stage

  Filed November 2024 and in examination; the claim read here is an original claim as filed and has not been allowed.

  The claim as amended, showing changes

  A computer-implemented method for generating Artificial Intelligence (AI) content for insurance claim data for processing an insurance claim in an insurance claims computer application, comprising the steps:

  1. capturing insurance claim data, in a computer processor, from at least one data storage component, via a computer network
  2. reformatting, in the computer processor, the captured insurance claim data in a confiauration suitable for analysis
  3. analyzing, in the computer processor, the captured insurance claim data by one or more Altechniques, to generate one or more prompts , wherein the analyzingstep automatically identifies, via one or moregenerative Altechniques, key data from the reformatted insurance claim data , wherein the one or more prompts are dynamically generated based on the analysis of the captured insurance claim data to select appropriate Almodels from a plurality of available Almodels for processing the specific type of insurance claim data captured, and wherein the one or more prompts indicate one or more Almodels to be used in conjunction with captured insurance claim data
  4. digitizing, in the computer processor, the captured insurance claim data so as to be processed by at least one generative Alsystem
  5. providing, from the computer processor, the digitized insurance claims data and associated one or more prompts to the at least one generative Alsystem, via an application programming interface (API) call, to cause the at least one generative Alsystem, using the one or more Almodels, to generate Alcontent responsive to the generated one or more prompts and associated captured insurance claims dat [ and]][
  6. ]][a[, wherein the generated Alcontent consists of one or more of: an insurance content summarization, a document summarization, claims queries, and classification of medical or biomedical information, each associated with the insurance claim
  7. providing, from the computer processor, the generated Alcontent to the insurance claim computer application [.]][, wherein the generated Alcontent causes the insurance claims computer application to automatically advance administration of the insurance claim. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 10/27/2025, most recent 3/27/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: US 11,170,450, Gunjan et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260141454A1 · filed 11/15/2024
• State Farm awaiting examination published 5/21/2026

  Chatbot to assist in vehicle shopping

  Describes a conversational agent that assists a user shopping for a vehicle, built on a machine-learned model trained by a forward-forward method that passes positive then negative data through each layer and adjusts its weights to tune a per-layer goodness metric.

  How it works

  Opens a session through a conversational AI agent, takes input that the user is considering buying a vehicle, computes a personalized insurance cost from information about the user and the vehicle, derives ownership information from the vehicle type and that insurance cost, and presents the ownership information back to the user.

  Where it fits

  Vehicle-shopping assistance that folds a personalized insurance-cost estimate into the ownership picture, supporting auto insurance quoting.

  Stage

  Filed January 2026, published May 2026 and awaiting examination; the claim read here is the original as filed, the broadest version the applicant will hold.

  The claim as filed

  A computer-implemented method for generating recommendations regarding vehicle purchases via a conversational artificial intelligence (AI) agent implemented by one or more processors, the method comprising:

  1. initiating, via the conversational AI agent, a session associated with a user of a computing device
  2. receiving, via the conversational AI agent, input indicating a potential purchase of a vehicle by the user
  3. determining, via the conversational AI agent, a personalized insurance cost for the user insuring the vehicle based upon information regarding the user and vehicle information regarding the vehicle
  4. determining, via the conversational AI agent, information regarding ownership of the vehicle based upon a type of the vehicle and the personalized insurance cost
  5. causing, via the conversational AI agent, the information regarding ownership of the vehicle to be presented to the user via the computing device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260141435A1 · filed 1/14/2026
• TIAA final rejection published 5/21/2026

  Systems and methods for technical debt risk management using artificial intelligence

  Describes classifying technical debt across an organisation and predicting its risk severity from internal, third-party and operational risk inputs.

  How it works

  Gathers organizational, third-party tool, operational-risk, and prior-assessment data, processes it through neural networks, generative AI models, and reinforcement learning to estimate technical-debt risk severity and likelihood, predicts a potential incident, adjusts network weights from evaluation metrics, and outputs the risk information.

  Where it fits

  Enterprise software and operational risk management; the claim describes internal IT governance rather than an insurance workflow.

  Stage

  Filed November 2024 and now under a final rejection; the claim read here was amended during prosecution, meaning the applicant narrowed it, not that the examiner agreed.

  The claim as amended, showing changes

  A computing system for detecting, classifying, and predicting risk severity and/or likelihood from technical debt within an organization, the computing system comprising:

  1. a memory having stored thereon computer-executable instructions that, when executed, cause the computing system to: obtain input data including internal organization information, third-party tool information, operational risk management information, previous technical debt assessments, previous asset rationalizations, and business capability information
  2. process the input data using a combination of one or more neural networks, generative artificial intelligence models, and reinforcement learning with human feedback
  3. generate the risk severity and/or likelihood from technical debt based on the processing
  4. predict, using the one or more neural networks, a potential incident from the technical debt based upon the risk severity and / or likelihood
  5. adj ust , based on one or more evaluation metrics associated with the predicted potential incident, one or more weights of the one or more neural networks
  6. and generate an output of risk information associated with the risk severity , the potential incident, and/or likelihood from technical debt. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 1/20/2026, most recent 8/13/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260141328A1 · filed 11/19/2024
• State Farm responded published 5/21/2026

  Systems and methods for water damage claims triage portal with computer vision

  Seeks to score a water damage claim from its text and images, then match it to the best-fit available adjuster by skill and availability, aimed at routing claims at intake rather than by manual triage.

  How it works

  runs a claim's text through a convolutional network to pull location, damage type and severity, its photos through a recurrent network that tracks damage across a series of images, fuses both into claim ratings, then scores each available adjuster on skill and availability and matches the claim to the best fit.

  Where it fits

  property claims intake and triage, auto-routing a water-damage claim to a suitable handler.

  Stage

  filed November 2024; in examination with claim 1 already amended, so the applicant has narrowed it during prosecution, not had it allowed.

  The claim as amended, showing changes

  A computer-implemented method for determining a best fit claims representative user, the computer-implemented method comprising:

  1. receiving, by one or more processors, text data and image data corresponding to a claim instance from a user device
  2. inputting, by the one or more processors, the text data into one or more trained machine-learning models comprisinqa convolutional neural network to determine a text vector corresponding to the text data, wherein the convolutional neural network segments the text data to extract one or more context items including location data, a damage type, and a damage severity
  3. inputting, by the one or more processors, the image data into the one or more trained machine-learning models comprising a recurrent neural network to determine an image vector corresponding to the image data, wherein the recurrent neural network processes a series of images received over a period of time to determine potential damage over the period of time and to identify one or more image features including a location of damage and a size of damage
  4. inputting, by the one or more processors, the text vector and the image vector into the one or more machine-learning modelsa multi-modal machine-learning model to determine one or more claim ratings corresponding to the claim instance, wherein the multi-modal machine-learning model combines the text vector and the image vector to determine one or more associations between the one or more context items extracted from the text data and the one or more image features identified in the image data
  5. receiving, by the one or more processors, claims representative user data corresponding to one or more claims representative users from one or more data stores, wherein the claims representative user data includes a claims representative user identifier, a claims representative user availability, and a claims representative user skill level
  6. determining, by the one or more processors, a claims representative user score for each of the one or more claims representative users based on the claims representative user data
  7. analyzing, by the one or more processors, the claims representative user data and the one or more claim ratings to determine theselecting, by the one or more processors, a best fit claims representative user[[,]] by comparing the one or more claim ratings to the claims representative user score for each of the one or more claims representative users to match the claim instance with the best fit claims representative user of the one or more claims representative user

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 3/30/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Basuri et al, Dorai et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260141322A1 · filed 11/19/2024
• CSAA awaiting examination published 5/21/2026

  Systems and Methods for Vehicle Navigation

  Proposes building an accident risk model from vehicle sensor and environmental data for a region, then answering navigation requests against that risk surface rather than travel time alone.

  How it works

  a modeling server collects sensor data from many vehicles plus environmental data for a region, trains an accident-risk model on it, and for a requested trip scores the attributes of several candidate routes by that risk; a mobile device then displays the routes, their risk attributes and a prompt to pick a better time to travel.

  Where it fits

  auto risk assessment and telematics-style routing, though the claim frames it as consumer trip navigation.

  Stage

  filed November 2025; this is a new claim substituted after the original claims were cancelled, typically following a rejection, and it remains in examination.

  The claim new claim

  A navigation system configured for use with a vehicle, the navigation system comprising:

  1. a modeling computing device, wherein the modeling computing device comprises a first processor and a first non-transitory computer-readable medium, having stored thereon first program instructions that, upon execution by the first processor, cause the modeling computing device to perform a first set of operations comprising: collecting sensor data from sensors attached to a plurality of vehicles operating within a geographic region and environmental data for the geographic region
  2. generating an accident risk model using one or more machine learning models, wherein the one or more machine learning models are configured to generate the accident risk model using the collected sensor data and environmental data
  3. receiving a request for navigating between a first geographic position and a second geographic position
  4. based on the received request, identifying attributes of a plurality of routes between the first geographic position and the second geographic position, wherein the identified attributes of the plurality of routes are based on at least the generated accident risk model
  5. a mobile computing device, wherein the mobile computing device comprises a second processor and a second non-transitory computer-readable medium, having stored thereon second program instructions that, upon execution by the second processor, cause the mobile computing device to perform a second set of operations comprising: receiving, from the modeling computing device, the identified attributes of the plurality of routes
  6. displaying, via a user interface of the mobile computing device, a graphical indication of : (i) the plurality of routes
  7. (ii) the identified attributes of the plurality of routes
  8. displaying, via the user interface of the mobile computing device, a suggestion prompt to choose a particular time to navigate at least one route of the plurality of routes. New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260139959A1 · filed 11/19/2025
• State Farm awaiting examination published 5/14/2026

  Method and system for virtual area visualization

  Describes assembling images of a region into a virtual model and presenting it as a navigable environment.

  How it works

  takes a virtual model of a region containing a structure, computes the structure's geographic coordinates, checks them against the coordinates of a customer's property, and on a match builds a 3D environment of the structure and embeds a link from it to that customer's digital record.

  Where it fits

  property underwriting and servicing, connecting a building's 3D view to the matching policyholder file.

  Stage

  filed January 2026; the claim is as filed, the original and broadest version, and sits early in examination before any office action.

  The claim as filed

  A method, comprising:

  1. receiving, by a processor, a virtual model of a geographic region, the geographic region including a first structure
  2. determining, by the processor and using the virtual model, first geographic coordinates of the first structure
  3. determining, by the processor, a match between the first geographic coordinates and second geographic coordinates of a second structure associated with a customer
  4. based on determining the match: generating, by the processor and based on the virtual model, a three-dimensional (3D) virtual environment illustrating the first structure disposed within the geographic region, generating, by the processor, a digital link between the first structure in the 3D virtual environment and a digital record associated with the customer, and providing, by the processor and to an electronic device, the 3D virtual environment including a representation of the digital link. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260134679A1 · filed 1/12/2026
• Clara Analytics awaiting examination published 5/14/2026

  Provider performance scoring using supervised and unsupervised learning

  Seeks to score a medical provider's predicted performance from the claims they have been involved in, combining supervised and unsupervised models, aimed at provider selection in claim management.

  How it works

  feature-engineers historical claims, clusters them into candidate models, then for a new claim runs one model to place the claim in a cluster and a second to place the provider in a cluster, and combines the two into a provider score standing for the provider's predicted performance.

  Where it fits

  claims handling, rating or selecting the provider attached to a claim by predicted performance.

  Stage

  filed October 2025; this is a new claim filed to replace cancelled original claims, usually after a rejection, and remains under examination.

  The claim new claim

  A computer-implemented method of generating accurate provider performance predictions, comprising:

  1. receiving, via a feature engineering engine, historical claim data from a historical claim database
  2. generating, via the feature engineering engine, feature engineered data based on the historical claim data
  3. clustering, via a clustering algorithm, the feature engineered data into a plurality of candidate models
  4. receiving, via at least one of the plurality of candidate models, new claim data from a client device
  5. generating, via a first candidate model of the plurality of candidate models, a claim cluster determination based on the claim data
  6. generating, via a second candidate model of the plurality of candidate models, a provider cluster determination based on the new claim data
  7. combining, via a provider recommendation tool, the claim cluster determination and the provider cluster determination
  8. generating, via a provider recommendation tool, a provider score based on the combination of the claim cluster determination and the provider cluster determination, wherein the provider score corresponds to a predicted performance of a provider. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260134312A1 · filed 10/20/2025
• Zesty.ai allowed published 5/14/2026

  Hail Frequency Predictions Using Artificial Intelligence

  Seeks to predict hail frequency at a property by contrasting locations that sustained hail damage with those that did not, aimed at pricing and underwriting hail exposure at address level.

  How it works

  trains one model on historical hail events, including damaging and non-damaging instances, to estimate the frequency of an event exceeding a damaging threshold, and a second on property features to estimate severity, then combines them with a target property's features to predict its damage extent and trigger an action.

  Where it fits

  property catastrophe pricing and underwriting, sizing hail exposure at a specific location.

  Stage

  filed December 2025; claim 1 is currently amended, so the applicant has already narrowed it during prosecution rather than having it allowed.

  The claim as amended, showing changes

  A computer implemented method comprising:

  1. obtaining historical data related to one or more climate events corresponding to one or more geographical areas, the one or more climate events including a first climate event, wherein the historical data describes multiple instances of the first climate event including dama ging instances and non-dama ging instances of the first climate event
  2. training, using one or more processors, a first climate event model based on the historical data
  3. wherein the first climate event model is trained to identify a climate event frequency indicative of a probability of an instance of the first climate event associated with a measurable characteristic that satisfies a dama ging threshold occurring
  4. training, using the one or more processors, a second climate event model based on the historical data and property feature data describing a plurality of properties represented in the one or more geographic areas represented in the historical data , the plurality of properties havingexperienced at least one damaging instance of the first climate event
  5. wherein the second climate event model is trained to identify a climate event severity associated with the first climate event
  6. receiving, using the one or more processors, one or more target locations
  7. obtaining, using the one or more processors, target property feature data describing one or more properties associated with the one or more target locations
  8. determining, using the one or more processors, a predicted extent of damage attributable to the first climate event in association with the one or more target locations, the extent of damage based on the features specific to the one or more properties associated with the one or more target locations, the predicted extent of damage determined using the first climate event model and the second climate event model
  9. taking, based on the predicted extent of damage attributable to the first climate event in association with the one or more target locations, one or more actions. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 4/1/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing

  References cited against it: Dhuvur et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260133342A1 · filed 12/22/2025
• State Farm responded published 5/14/2026

  Flat tire detection system implementing acceleration data

  Describes a sensor system that reads vehicle acceleration data to classify a condition such as a flat tire, then queries a coverage database by vehicle identifier to select a location-aware response action, oriented to roadside servicing and claims handling.

  How it works

  maps a vehicle's acceleration data to a set of predefined conditions and driving behaviors; when the condition is a flat tire, it queries a database keyed by the vehicle's identifier for its service coverage, then picks response actions from that coverage and the vehicle's location and triggers one.

  Where it fits

  auto telematics and servicing, matching a detected fault to the vehicle's roadside or service coverage.

  Stage

  filed November 2024; claim 1 is currently amended, meaning the applicant has narrowed it in prosecution, not that it has been allowed.

  The claim as amended, showing changes

  A system, comprising:

  1. one or more processors
  2. one or more non-transitory, computer-readable media including instructions which, when executed by the one or more processors, cause the one or more processors to: obtain, using one or more sensors of a sensor device, acceleration data associated with a vehicle
  3. determine, based on the acceleration data, a vehicle condition of the vehicle from among [[of]] a plurality of predefined vehicleconditions including vehicle conditions and driving behaviors by mapping the acceleration data to the plurality of predefined conditions
  4. in response to the determined vehicle condition including a flat tire, query a database which includes a plurality of vehicle identifiers associated with service coverage, wherein querying the database comprises using a query including a vehicle [[an]] identifier of the vehicle to determine service coverage for the vehicle
  5. determine one or more response actions based on the service coverage for the vehicle and a location of the vehicle
  6. cause implementation of at least one of the one or more response actions. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 4/30/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: US 11,673,579.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260131802A1 · filed 11/12/2024
• Allstate under examination published 5/14/2026

  Kinetic insights machine

  Proposes identifying vehicle faults from vibration data captured off a component, aimed at detecting mechanical problems before they present as a claim.

  How it works

  feeds vibration data from a vehicle component into machine-learning models to identify a vehicle problem, then sends a service facility's device commands to display a re-creation of the vehicle event, where that re-creation is built from the vibration data and the identified problem.

  Where it fits

  auto diagnostics and claims servicing, reconstructing a component fault for a repair or service facility.

  Stage

  filed October 2025; this is a new claim presented in place of cancelled original claims, usually after a rejection, and is still in examination.

  The claim new claim

  A kinetic insights method comprising:

  1. at a computing device comprising at least one processor, a communication interface, and memory: receiving vibration data corresponding to a vehicle component
  2. identifying, using one or more machine learning models and based on the vibration data, a vehicle problem corresponding to the vibration data
  3. transmitting, after identifying the vehicle problem and to an enterprise user device corresponding to a service facility, one or more first commands directing the enterprise user device to display a user interface indicating a re-creation of a vehicle event
  4. displaying, via the enterprise user device, the user interface indicating the re- creation of the vehicle event, wherein the re-creation of the vehicle event is determined based on vibration data and the vehicle problem. New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/10/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  References cited against it: Son et al, Zhang et al, Brannan et al, Johnson et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260131798A1 · filed 10/6/2025
• Hartford awaiting examination published 5/7/2026

  Customized risk relationship user interface workflow

  Describes pulling third-party data on a prospective customer and driving an underwriting workflow interface from what that data returns.

  How it works

  takes a new prospect's parameters, pulls third-party data and reads internal cloud records through a stored procedure that evolves the schema into an incremental view, then runs a machine-learning step over all of it to turn a first question workflow into a shorter second one that removes collection screens, cuts questions and prefills fields.

  Where it fits

  underwriting intake and onboarding, tailoring the application questionnaire to each prospect.

  Stage

  filed January 2026; the claim is as filed, its original and broadest form, and sits early in examination before any office action.

  The claim as filed

  A user interface workflow customization system implemented via a back-end application computer server, comprising:

  1. (a) a user information data store that contains electronic records associated with users, each electronic record including an electronic record identifier and user parameters
  2. (b) the back-end application computer server, associated with the enterprise and coupled to the user information data store, including: a computer processor, and a computer memory, coupled to the computer processor, storing instructions that, when executed by the computer processor, cause the back-end application computer server to: receive, from a remote user device, information about a new potential risk relationship customer of an enterprise, including at least one new user parameter, based on the new user parameter, access third-party data about the new potential risk relationship customer, utilize a stored procedure of a cloud computing environment curation engine to read data about the new potential risk relationship customer from an internal table of cloud data, process the data read from the internal table to dynamically evolve a schema and create an incremental view of cloud data, use the created incremental view to read and output a current batch of cloud data about the new potential risk relationship customer, provide a first user interface workflow associated with all of: an order of questions on a user interface, a selection of questions on the user interface, and an online-to-offline handoff process
  3. execute a machine learning algorithm that processes the information about the new potential risk relationship customer, the third-party data, and the current batch of cloud data, wherein an output of the execution is a second user interface workflow, and based on the execution and the third-party data: (i) data collection displays that receive information from the new potential risk relationship customer and are included in the first user interface workflow are eliminated from the second user interface workflow, (ii) a number of questions included on presented data displays included in the first user interface workflow is reduced on the second user interface workflow, and (iii) data elements are prefilled on the user interface associated with the second user interface workflow, wherein avoiding transmission of both the eliminated displays and questions reduces el

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260127678A1 · filed 1/2/2026
• State Farm awaiting examination published 5/7/2026

  Systems and methods for AI based recommendations for object placement in a home

  Describes training on existing room layouts, then recommending where objects should sit in a room from its LIDAR dimensions.

  How it works

  builds room dimensions by measuring LIDAR data from a LIDAR camera, feeds those dimensions plus an object's attributes including its color into a machine-learning model to recommend a placement, then shows the user's own placement alongside the model's so the two can be compared.

  Where it fits

  general-purpose home interior visualization; the claim states no specific insurance workflow.

  Stage

  filed January 2026; the claim is as filed, the original broadest version, and is early in examination with no office action yet.

  The claim as filed

  A computer-implemented method for object placement based upon light detection and ranging (LIDAR) data and machine learning, the computer-implemented method comprising:

  1. generating, via one or more processors, room data comprising a plurality of dimensions of a room by: receiving, via the one or more processors, LIDAR data generated from a LIDAR camera
  2. measuring, via the one or more processors, the plurality of dimensions of the room based upon analysis of the LIDAR data
  3. with a machine learning algorithm, generating, via the one or more processors, a recommendation for placement of the object in the room based upon: (i) the generated room data, and (ii) object data including color data of the object
  4. receiving, via the one or more processors, an object placement in the room from a user
  5. displaying, via the one or more processors, both: (i) a representation of the object placement in the room from the user, and (ii) a representation of the object placement generated by the machine learning algorithm, thereby allowing the user to compare the placements. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260127433A1 · filed 1/2/2026
• State Farm under examination published 4/30/2026

  Generative machine learning models for generating roof damage images

  Describes a generative diffusion model that turns text describing roof attributes and damage into synthetic images of damaged roofs, proposed to augment training data for roof damage detection models used in property claims and assessment.

  How it works

  takes text describing a roof surface or damage attribute, feeds it to a generative model trained to produce images of damaged roof surfaces, and outputs a synthetic image of a damaged roof, intended to expand scarce training data for damage-detection models.

  Where it fits

  property claims tooling, manufacturing training images for roof-damage assessment models.

  Stage

  filed October 2024; the claim is as filed, its original and broadest form, and remains in examination.

  The claim as filed

  A method for generating synthetic roof images, the method comprising:

  1. receiving text data indicating at least one of a roof surface attribute or a roof damage attribute
  2. providing the text data as input to a generative model, wherein the generative model is a machine learning (ML) model trained to generate images of damaged roof surfaces
  3. generating, based on an output of the generative model, a synthetic image of a damaged roof. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/19/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260120353A1 · filed 10/31/2024
• State Farm under examination published 4/30/2026

  Systems and methods for a scalable and coordinated enterprise test data management system

  Describes an enterprise test data management method that maps a data asset, splits a workflow into portioned parallel sub-workflows keyed to sensitive data subsets, and aggregates the results back to the requester.

  How it works

  on a request naming a data asset and a sensitive subset, a machine-learning model maps where that sensitive data sits, the system splits the asset into an optimized number of parallel workflows, runs them while obfuscating the sensitive fields, and aggregates the outputs back to the requester.

  Where it fits

  general enterprise test-data and data-privacy handling; the claim names no insurance workflow.

  Stage

  filed October 2024; the claim is as filed, the original broadest version, and sits in examination.

  The claim as filed

  A computer-implemented method for test data management, the computer-implemented method for test data management comprising:

  1. receiving, by one or more processors, a request to execute a workflow from a device, wherein the request includes a data asset identifier corresponding to a data asset, a sensitive data subset, a sensitive data subset type, and a user identifier
  2. generating, by the one or more processors and a machine-learning model, a data asset map corresponding to the data asset identifier, wherein the data asset map identifies a location of the sensitive data subset and the corresponding sensitive data subset type within the data asset
  3. analyzing, by the one or more processors, the data asset and the data asset map to determine an optimized number of one or more portioned workflows for parallel processing
  4. parsing, by the one or more processors, the data asset based on the optimized number of one or more portioned workflows of the workflow for parallel processing
  5. executing, by the one or more processors, the one or more parsed workflows optimized for parallel processing, wherein the executing includes obfuscating the sensitive data subset
  6. aggregating, by the one or more processors, the data from the one or more parsed workflows into an aggregated dataset
  7. transmitting, by the one or more processors, the aggregated dataset to a device corresponding to the user identifier. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 9/3/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260119379A1 · filed 10/31/2024
• Zesty.ai allowed published 4/30/2026

  Hail Frequency Predictions Using Artificial Intelligence

  Describes training and validating a hail damage frequency model from a set of properties labelled by whether they experienced hail damage, using each property's location and features, which would serve property risk assessment.

  How it works

  predicts hail damage extent at a target location from its property features using two pre-trained models, one estimating event frequency above a damaging threshold from historical damaging and non-damaging instances, the other estimating severity from features of previously damaged properties, then takes an action on the result.

  Where it fits

  property catastrophe pricing and underwriting, sizing hail exposure for specific locations.

  Stage

  filed December 2025; claim 1 is currently amended, so the applicant has already narrowed it in prosecution rather than having it allowed.

  The claim as amended, showing changes

  A computer implemented method comprising:

  1. determining, using one or more processors, a predicted extent of damage attributable to a first climate event in association with one or more target locations, the predicted extent of damage based on property feature data specific to the one or more properties associated with the one or more target locations, the predicted extent of damage determined using a first climate event model and a second climate event model
  2. wherein the first climate event model has been trained based on historical data related to at least the first climate event corresponding to one or more geographical areas , wherein the historical data describes multiple instances of the first climate event including dama ging instances and non-dama ging instances of the first climate event
  3. wherein the first climate event model has been trained to identify a climate event frequency indicative of a probability of an instance of the first climate event associated with a measurable characteristic that satisfies a dama ging threshold occurring
  4. wherein the second climate event model has been trained based on historical data related to a first set of property features associated with a plurality of properties represented in one or more geographic areas, the plurality of properties having experienced at least one dama ging instance of the first climate event
  5. wherein the second climate event model has been trained to identify a climate event severity associated with the first climate event
  6. taking, based on the predicted extent of damage attributable to the first climate event in association with the one or more target locations, one or more actions. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 4/1/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing

  References cited against it: Dhuvur et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260118550A1 · filed 12/24/2025
• State Farm responded published 4/30/2026

  Combined segmentation and computer vision machine learning models for roof damage detection

  Describes pairing a segmentation model with a computer vision model, including a recurrent network across multiple images, to assess roof damage from imagery and output damage cause, severity and possible fraud signals for property claims.

  How it works

  runs a roof image through a segmentation model to get a mask, folds that mask into the image as extra channels carrying segment shape, size or depth labels, then feeds the enriched multichannel image to a second computer-vision model that renders the roof-damage determination.

  Where it fits

  property claims, assessing roof damage from imagery.

  Stage

  filed October 2024; claim 1 is currently amended, meaning the applicant has narrowed it during prosecution, not that it has been allowed.

  The claim as amended, showing changes

  A method for machine learning detection of roof damage, the method comprising:

  1. receiving image data representing a roof surface
  2. providing the image data as input to a first segmentation machine learning model
  3. determining a segmentation mask based on an output of the first segmentation machine learning model
  4. modifying the image data, into a modified image representation of the roof surface, comprising one or more additional data channels associated with based onthe segmentation mask , the additional data channels comprising at least one of a segment shape label, a segment size label, or a segment depth label
  5. providing the modified image representation as input to a second computer vision machine learning model
  6. performing a roof damage determination based on a second output of the second computer vision machine learning model. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 5/8/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Frei et al, Frei.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260118280A1 · filed 10/31/2024
• State Farm awaiting examination published 4/23/2026

  Systems and methods for IoT device modeling and interfacing

  Describes using AI models to interface with home IoT devices by inferring each device's protocol, command or data format, storing a device dataset, and answering natural language queries with retrieval augmented generation.

  How it works

  reads a home IoT device's transmission to infer its protocol, command or data format, stores that as a device dataset, then on a natural-language request uses retrieval-augmented generation over the dataset to produce interfacing commands, queries the device for its status or configuration, and answers in natural language.

  Where it fits

  smart-home and property monitoring, giving a natural-language interface to household IoT devices.

  Stage

  filed January 2025; the claim is as filed, its original and broadest form, and remains under examination.

  The claim as filed

  A system for interfacing with a plurality of Internet of Things (IoT) devices within a local network of a residential building using one or more artificial intelligence (AI) models, comprising:

  1. one or more memory devices having instructions stored thereon that, when executed by one or more processors, cause the one or more processors to perform operations comprising: receiving at least one transmission of an IoT device of a plurality of IoT devices connected within the local network
  2. determining, using the at least one transmission, at least one of a protocol, a command, or a data format of the IoT device
  3. generating and storing a device dataset comprising the at least one of the protocol, the command, or the data format of the IoT device
  4. receiving, via a user interface, a natural language input corresponding to the IoT device
  5. generating a natural language response to the natural language input using retrieval-augmented generation (RAG) by: applying the device dataset and the natural language input as input to the one or more AI models to cause the one or more AI models to generate an output comprising one or more interfacing commands for initiating communications with the IoT device based upon at least one of the protocol, the command, or the data format of the IoT device
  6. initiating a communication with the IoT device using the one or more interfacing commands to obtain a status or configuration of the IoT device
  7. generating the natural language response using the status or configuration of the IoT device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260113249A1 · filed 1/10/2025
• State Farm awaiting examination published 4/23/2026

  Sensing peripheral heuristic evidence, reinforcement, and engagement system

  Describes home sensors, including monitoring of device electricity use, whose data a processor analyzes for anomalies to infer a condition such as a medical one and notify a caregiver, oriented to in-home monitoring.

  How it works

  captures data from home sensors and runs it through a model trained on that home's data to identify an individual's patterns, then periodically sends a caregiver's device a snapshot report of those patterns and how they have changed from the prior report.

  Where it fits

  health and life servicing, monitoring an individual at home for a caregiver, such as aging in place.

  Stage

  filed December 2025; the claim is as filed, the original broadest version, and is early in examination before any office action.

  The claim as filed

  A computer-implemented method for identifying patterns associated with an individual in a home environment and periodically generating and transmitting indications of the patterns to a device associated with a caregiver of the individual, comprising:

  1. capturing data detected by a plurality of sensors associated with the home environment
  2. analyzing, by one or more processors, the captured data using a machine learning model trained using a dataset associated with the home environment, to identify one or more patterns associated with an individual in the home environment
  3. periodically generating and transmitting, by the one or more processors, to a device associated with a caregiver of the individual, notifications indicating the one or more patterns associated with the individual, wherein each notification comprises a snapshot report generated periodically and the snapshot report includes an indication of the one or more patterns associated with the individual and a change from a prior snapshot report. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260112259A1 · filed 12/12/2025
• Allstate under examination published 4/23/2026

  Chatbot system and machine learning modules for query analysis and interface generation

  Describes training a model per chatbot plus a routing model that sends each query to the right one, aimed at directing customer questions to the assistant best suited to answer.

  How it works

  parses a query to identify a category and two sub-categories, routes different portions of it to at least two specialized chatbot models chosen by those sub-categories, generates a combined response from them within the conversation, and trains the selected models by tying the response back to their chatbots.

  Where it fits

  customer-service and servicing chat, splitting a query across specialized assistants; the claim is not insurance-specific.

  Stage

  filed June 2025; this is a new claim substituted for cancelled original claims, usually after a rejection, and remains in examination.

  The claim new claim

  A chatbot system configured to analyze query data, the chatbot system comprising:

  1. a plurality of machine learning models and a plurality of chatbots, each of the plurality of machine learning models corresponding to a respective chatbot of the plurality of chatbots
  2. a computing platform comprising a processor, a non-transitory computer-readable memory communicatively coupled to the processor, and machine-readable instructions stored in the memory that, when executed by the processor, cause the processor of the computing platform to, across at least one or more intelligent platforms, technical platforms, or combinations thereof: parse a query comprising a content to identify at least a category and at least a first sub- category and a second sub-category, each associated with the category corresponding to the query
  3. route the query to at least two selected models of the plurality of machine learning models based on the content of the query and the at least the first sub-category and the second sub-category that are identified such that the query is routed to at least two selected chatbots of the plurality of chatbots, wherein at least a first portion of the query is routed to a first selected chatbot of the at least two selected chatbots based on the first sub-category as identified and at least a second portion of the query is routed to the a second selected chatbot different from the first selected chatbot based on the second sub-category as identified, wherein the first portion is the same as, partially the same as, or different from the second portion of the query
  4. generate a response to the query, using the at least two selected models and corresponding at least two selected chatbots, as part of a conversation between a user of a client device and the computing platform
  5. train the at least two selected models of the plurality of machine learning models by generating an association between the response and the at least two selected chatbots. New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 6/23/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260111968A1 · filed 6/27/2025
• State Farm awaiting examination published 4/23/2026

  Systems and methods for enhancing waste disposal and energy efficiency using sensor and alternative power technologies

  Describes reading waste and recycling sensor data near a bin to determine fill level, aimed at waste handling and energy efficiency in a monitored property.

  How it works

  collects energy data from a user's electric devices, builds an energy-consumption profile of their patterns, picks a monitoring plan by matching that profile to a reference pattern with threshold values, then compares later energy readings to those thresholds and alerts the user's device when a device's operational-health status changes.

  Where it fits

  home and property monitoring, flagging appliance or device health from energy use; the claim is not insurance-specific.

  Stage

  filed December 2025; the claim is as filed, its original and broadest form, and remains in examination.

  The claim as filed

  1. A computer system for monitoring and generating real-time alerts relating to an operational health of one or more electric devices associated with a user, the computer system comprising at least one sensor, at least one memory device, and at least one processor in communication with the at least one sensor and the at least one memory device, the at least one processor programmed to: collect, via the at least one sensor, energy data from the one or more electric devices
  2. build, using the collected energy data, an energy consumption profile for the user, wherein the energy consumption profile includes energy consumption patterns of the one or more electric devices
  3. select, based upon the energy consumption profile, an energy monitoring plan by comparing the energy consumption profile to a reference pattern of energy consumption associated with the energy monitoring plan, the energy monitoring plan including one or more threshold values
  4. collect additional energy data for the one or more electric devices over a subsequent period of time
  5. compare the collected additional energy data to the one or more threshold values of the selected energy monitoring plan
  6. in response to the comparison indicating that a status of the operational health of the one or more electric devices has changed, cause an alert to be displayed on one or more client devices associated with the user indicating that the status of the operational health of the one or more electric devices has changed. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260111855A1 · filed 12/17/2025
• Swiss Re awaiting examination published 4/23/2026

  System for automated detection and assessment of employee-based electronic links of a unit to external actors, and electronic method thereof

  Seeks to detect and weight electronic links between an employee and external actors through interaction data mining, aimed at assessing exposure arising from those connections.

  How it works

  capturing devices tied to an employee's unit collect the employee's account parameters and their interaction data; an accumulation module extracts the external actors, then a weighting unit with a machine-learning component measures each interaction and assigns weighting factors that score the strength of each employee-to-actor link.

  Where it fits

  insider and connection risk assessment; the claim frames it broadly, not as a specific insurance line.

  Stage

  filed September 2025; the claim is as filed, its original and broadest form, and sits in examination.

  The claim as filed

  An electronic system for automated detection and assessment of employee-based electronic links of at least one employee to external actors electronically interacting with the at least one employee wherein the electronic system is based on automated interaction data mining and impact weighting of the employee-based electronic links, the electronic system comprising:

  An electronic system for automated detection and assessment of employee-based electronic links of at least one employee to external actors electronically interacting with the at least one employee wherein the electronic system is based on automated interaction data mining and impact weighting of the employee-based electronic links, the electronic system comprising: processing circuitry configured to implement a trigger module with a plurality of capturing devices connectable to at least one employee's unit for automatically capturing employee data comprising employee-based parameter values for one or more employee-based parameters identifying at least one electronic employee account associated to the at least one employee or employee's unit and for capturing link data comprising interaction parameter values for one or more interaction parameters characterizing electronic interactions of the employee-based electronic links, a trigger table providing an electronic link network based on selected employee parameter values or selected interaction parameter values, and a link network module for data mining and transfer control connected to the trigger module and the trigger table, wherein the link network module at least comprises or is connectable to an accumulation device connected to the trigger module automatically receiving the employee data and the link data and to a repository unit, the accumulation device comprising a data extraction algorithm extracting external actor data comprising actor parameter values of actor parameters included in the link data, and storing the employee data, the link data and the external actor data in the repository unit, and a weighting unit connected to the accumulation device or the repository unit and comprising a weighting algorithm automatically assigning weighting factors to one or more of the captured interaction parameter values or to the employee-based electronic links, and providing weighted link data for the employee-based electronic links, wherein the link strength and the weighting factor are determined by measuring the employee and interaction parameters and capturing additional external link parameter data related to the interactions and links, and the weighting algorithm comprising a machine learning structure assessing the link strength contribution of a data exchange interaction based on the measured employee a

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260111828A1 · filed 9/15/2025
• Elevance under examination published 4/23/2026

  Personalized Smart Provider Search

  Describes matching a member's stated criteria and characteristics against provider records to recommend providers, aimed at steering members toward a suitable clinician.

  How it works

  precomputes a vector of a member's characteristics, and for a provider request retrieves candidate providers each as a vector carrying quality and cost factors, then scores each provider against the member vector and the member's criteria to return an ordered, member-specific provider list.

  Where it fits

  health provider search and network steerage, ranking providers by fit, cost and quality for a member.

  Stage

  filed October 2025; the claim is as filed, its original and broadest form, and remains under examination.

  The claim as filed

  A method for recommending one or more providers to a member, the method comprising:

  1. at a server: computing a member vector representative of one or more member characteristics representative of the member
  2. after computing the member vector, receiving, from a computing device, a provider request including a member identifier and member specified criteria, wherein the member specified criteria include member preference for a provider treating a condition of the member, and wherein the one or more member characteristics are separate and distinct from the member specified criteria
  3. after receiving the provider request: retrieving the member vector
  4. retrieving one or more provider identifiers for association with the member specified criteria
  5. for each provider identifier, retrieving one or more provider characteristics for association with the provider identifier, represented as a respective pre-computed provider vector, wherein the provider vector includes a provider quality factor and a provider cost factor
  6. analyzing the member vector, the member specified criteria, and the provider vector to generate an ordered list of the providers, wherein the ordered list of the providers is based upon a member-provider score for each provider, wherein the member-provider score is generated as a function of the member vector and the provider vector
  7. transmitting the ordered list of the providers to the computing device.

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/14/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260111427A1 · filed 10/3/2025
• TIAA responded published 4/16/2026

  Techniques for Artificial Intelligence-Based Security Governance of Computing System Interfaces

  Describes generating an execution script that exercises an interface function under several security policies, aimed at testing whether an externally callable interface stays compliant.

  How it works

  generates a script that exercises an API's externally invocable functions under several security policies, runs it to capture policy-execution outputs, evaluates how far the API complies with those policies, and either alerts when compliance falls below a threshold or logs the execution details otherwise.

  Where it fits

  general API security governance and compliance monitoring; the claim names no insurance workflow.

  Stage

  filed October 2024; claim 1 is currently amended, so the applicant has narrowed it in prosecution rather than having it allowed.

  The claim as amended, showing changes

  A computer-readable medium including instructions that, when executed on a processor, cause the processor to perform operations for monitoring security compliance in a computing system, the operations comprising:

  1. generating an execution script to execute at least one function of an application proQ ramminginterface (API) under a plurality of security policies, the at least one function being invoked by a device outside the computing system , and the interfaceAPI comprising executable code segments that, when executed, access information within at least a portion of the computing system
  2. executing the execution script to generate policy execution outputs
  3. evaluating an extent to which the interfaceAPI complies with the plurality of security policies associated with the computing system based on the policy execution outputs
  4. providing an alert if the extent to which the interfaceAPI complies with the plurality of security policies fails to meet a threshold , and accessing a database to log [[the]] a extent or details of the executing of the at least one function otherwise. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 5/7/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  The examiner has pointed at allowable subject matter in claims 6, 7, 11, 12, 19, 20. That is the examiner naming the limitation that would earn the patent, not a statement that it will be granted.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260106901A1 · filed 10/16/2024
• CVS Health responded published 4/16/2026

  Multimodal Recognition and Authentication of Pharmaceuticals

  Seeks to count and verify dispensed prescription product from camera images at the dispensing site, aimed at checking a fill against the prescription without a second physical count.

  How it works

  a receptacle holds a pharmaceutical while a camera, spectrometer, olfactory sensor and audio sensor each capture data; a processor extracts a feature from each and a machine-learning model fuses them into a recognition-confidence value for a pharmaceutical type, returning an indicator the user acts on.

  Where it fits

  pharmacy and health operations, identifying or verifying a dispensed pharmaceutical.

  Stage

  filed October 2024; claim 1 is currently amended, meaning the applicant has narrowed it during prosecution, not that it has been allowed.

  The claim as amended, showing changes

  A system, comprising:

  1. a receptacle configured SVG 18913448.07-30-2026.MSLKBSM0X11X148.CLM.1.4.929.738.982.775.svg 0.123 0.177 Chemistry Black and white receive a pharmaceutical product
  2. and insert into an enclosure
  3. a camera mounted to the enclosure and configured to capture image data for at least one image of the pharmaceutical product in the receptacle
  4. a spectrometer mounted to the enclosure and at least one sensorconfigured to capture spectral sensordata for at least one sensor signalbased on the pharmaceutical product in the receptacle
  5. an olfactory sensor mounted to the enclosure and configured to capture olfactory data based on response to low-concentration chemicals suspended in air in the receptacle
  6. an audio sensor mounted to the enclosure and configured to capture audio data based on response to vibration of the pharmaceutical product in the receptacle
  7. and at least one processor configured to: receive the image data, the spectral data, the olfactory data, and the audio sensordata
  8. determine at least one visual feature from the image data
  9. determine at least one spectral data sensorfeature from the spectral sensordata
  10. determine at least one olfactory data feature from the olfactory data
  11. determine at least one audio data feature from the audio data combine, using applyat least one machine learning model , [[to]] the at least one visual feature, the at least one spectral data feature, the at least one olfactory data feature, and the at least one audio data sensorfeature to determine a recognition confidence value for a pharmaceutical type
  12. return a recognition indicator based on the recognition confidence value for the pharmaceutical type, wherein, responsive to the recognition indicator, a user determines a disposition of the pharmaceutical product in the receptacle. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 5/5/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260106008A1 · filed 10/11/2024
• CVS Health awaiting examination published 4/16/2026

  Product identification with machine learning

  Describes determining a product's status and reconciling it against retrieved descriptions, aimed at identifying products where catalogue records disagree.

  How it works

  takes a selected product with its category and description, looks up its status in a database, retrieves descriptions of other products in the same category, and feeds those plus the product's description to a machine-learning model that outputs correlations, from which it recommends replacement products.

  Where it fits

  general product matching and replacement recommendation; the claim states no insurance workflow.

  Stage

  filed October 2025; the claim is as filed, its original and broadest form, and remains in examination.

  The claim as filed

  1. A system comprising one or more memory devices storing instructions thereon that, when executed by one or more processors, cause the one or more processors to: receive, via a user interface, a selection of a product, the product having a first category and a description
  2. determine, based on a query of a database, a status of the product
  3. retrieve, based on the status of the product, a plurality of descriptions of a plurality of products, the plurality of products having the first category
  4. provide, to a machine learning (ML) model, the plurality of descriptions and the description of the product to cause the ML model to generate one or more outputs that represent correlations between the product and one or more products of the plurality of products
  5. identify, based on the one or more outputs, the one or more products of the plurality of products
  6. provide, via the user interface, a recommendation to replace the product with the one or more products of the plurality of products. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260105509A1 · filed 10/9/2025
• CVS Health awaiting examination published 4/16/2026

  Product description generation with machine learning

  Describes retrieving product information, detecting where it contains proprietary content, and generating a description from what remains.

  How it works

  retrieves a product's information including its existing description, detects proprietary content in it, removes that content, then feeds the scrubbed information to a machine-learning model that generates a second description conforming to specified characteristics.

  Where it fits

  general automated content generation with proprietary-data removal; the claim names no insurance workflow.

  Stage

  filed October 2025; the claim is as filed, its original and broadest form, and remains under examination.

  The claim as filed

  1. A system comprising one or more memory devices storing instructions thereon that, when executed by one or more processors, cause the one or more processors to: retrieve information for a product of an entity, the information comprising a first description of the product
  2. determine, based on an evaluation of the information, that the information includes proprietary information associated with the product or the entity
  3. modify the information to remove the proprietary information from the information
  4. input the modified information into a machine learning model
  5. generate, using the machine learning model, a second description of the product using the modified information, the machine learning model configured to generate the second description to conform with one or more characteristics. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260105504A1 · filed 10/9/2025
• Equifax awaiting examination published 4/16/2026

  Artificial intelligence techniques for identifying identity manipulation

  Seeks to derive risk signals from an entity's own data and its interaction history to judge whether an identity has been manipulated, aimed at synthetic and altered identity fraud.

  How it works

  derives a risk signal for a target entity with an AI model, builds one graph of the entity's identity data and another of its interaction history, links them, applies the risk signal to the combined graph to compute a risk indicator, and blocks the entity from the environment when it signals malicious behavior.

  Where it fits

  fraud detection and access control, screening entities for identity manipulation.

  Stage

  filed December 2025; the claim is as filed, its original and broadest form, and sits early in examination.

  The claim as filed

  A system comprising:

  1. a non-transitory computer-readable medium comprising instructions that are executable by the processor to cause the processor to perform operations comprising: determining a risk signal associated with a target entity using an artificial intelligence model
  2. generating a first graph structure of a linked graph structure, wherein the first graph structure represents identity data about the target entity
  3. generating a second graph structure of the linked graph structure, wherein the second graph structure represents historical interaction data associated with the target entity
  4. linking the first graph structure and the second graph structure to form the linked graph structure, wherein an identity indicated by the first graph structure is associated with an interaction of the historical interaction data of the second graph structure
  5. applying the risk signal to the linked graph structure to determine a risk indicator associated with the target entity
  6. preventing, based on determining that the risk indicator indicates that the target entity is associated with malicious behavior, the target entity from accessing the interactive computing environment. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260105162A1 · filed 12/16/2025
• Allstate awaiting examination published 4/2/2026

  Commercial claim processing platform using machine learning to generate shared economy insights

  Describes training on historical claim processing records to produce commercial and shared-economy insights, aimed at exposures that sit outside conventional personal or commercial lines.

  How it works

  trains a primary model to tell a claim-processing request from a claim query, do similarity matching against historical claims for automated decisions, and pick specialist secondary models; at run time it either outputs a claim decision or routes sub-queries to the secondary models and returns their answers.

  Where it fits

  commercial claims processing and automated decisioning, including shared-economy exposures.

  Stage

  filed December 2025; the claim is as filed, its original and broadest form, and remains under examination.

  The claim as filed

  A computing platform comprising:

  1. at least one processor
  2. a communication interface communicatively coupled to the at least one processor
  3. memory storing computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: during a training phase: train one or more secondary machine learning models to provide corresponding answers to one or more sub queries of a claim related query
  4. train a primary machine learning model to perform operations comprising a) distinguishing between a claim processing request and a claim related query, b) performing similarity matching with historical claim data for automated claim decision-making, and c) identifying one or more secondary machine learning models for answering one or more sub queries of a claim related query
  5. during a live phase: receive, from a user computing device, a claim related request
  6. after determining, by the primary machine learning model, that the claim related request comprises a claim processing request, output, by the primary machine learning model, claim information, wherein the claim information comprises a claim decision associated with the claim processing request
  7. after determining, by the primary machine learning model, that the claim related request comprises a claim related query comprising one or more sub queries: determine one or more secondary machine learning models trained to answer the one or more sub queries
  8. output, by the determined one or more secondary machine learning models, claim information, wherein the claim information comprises one or more answers to the one or more sub queries
  9. after processing the claim related request, communicate, to the user computing device, the claim information associated with the claim related request. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260094215A1 · filed 12/8/2025
• State Farm under examination published 4/2/2026

  Systems and methods for model-based analysis of damage to a vehicle

  Describes a system that trains damage classification models on historical repair records, verifies that submitted photos of a damaged object are usable, then infers object type, damage type, repair amount and repair time to support claim estimating.

  How it works

  Trains damage classification models on historical damage and repair data, builds a digital view of a submitted object, checks each image is properly captured and asks the user to reshoot any that are not, then identifies object and damage type and runs selected models to output a repair amount and time.

  Where it fits

  Auto claim intake and repair estimating from policyholder photos.

  Stage

  Filed July 2023 and under examination; the claim read here was already narrowed by amendment, not the original.

  The claim as amended, showing changes

  A computer system for model-based analysis of damage to an object, the computer system comprising at least one processor in communication with at least one memory device, wherein the at least one processor is configured to:

  1. continuously train a plurality of damage classification models using historical damage data associated with damages and repairs for a plurality of objects continuously received by the computer system, wherein each of the plurality of damage classification models is configured to determine (i) an amount of damage to an object based upon a type of the object and a type of damage to the object, and (ii) how the damage would be repaired
  2. receive, from a user computing device associated with a user,i) a request for an estimate to repair a candidate object
  3. receive, from the user computing device,and (ii) a plurality of images , each capturing at least one portion of the candidate object to repair
  4. generate a digital view of the candidate object by comparing each of the received plurality of images to an orientation computer model associated with the candidate obj ect
  5. determine that each of the received plurality of images is properly captured by (i ) matching each received image to at least a portion of the generated digital view or (ii) determiningthat each received image satisfies an analysis threshold associated with imageacquisition parameters of each received imagein response to determining that one of the received plurality of images fails to be properly captured , (i ) generate instructions to recapture the at least one portion of the candidate object initiallycaptured in the one of the received plurality of images and (ii) cause the user computing device to display the instructions
  6. in response to determining that each of the received plurality of images is properly captured: determine, from the received plurality ofproperly captured images and using image recognition tools, the type of the object and the type of the damage to the candidate object
  7. select one or more of the plurality of trained damage classification models based upon the determined type of the object and the determined type of the damage to the candidate object
  8. input, into the selected one or more trained damage classification models, the determined type of the object and the determined type of the damage to the candidate object
  9. output, from the selected one or more trained damage classification models, [[an]] the amount of damage to

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 3 office actions

  First action 3/20/2025, most recent 6/8/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  A further rejection under section 103 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260094128A1 · filed 7/21/2023
• Hartford awaiting examination published 4/2/2026

  System and method for a generative artificial intelligence model gateway

  Describes a gateway that screens an incoming prompt for personally identifiable information and only when the prompt is PII-free routes it to a selected large language model and returns the output, governing enterprise use of generative AI.

  How it works

  Receives a prompt at an image or text component, runs a trained model to determine a PII status and returns a PII response, then only when the prompt is PII-free takes a selection of a large language model, checks that model's status, transmits the prompt to it and returns its output through an interface.

  Where it fits

  Governs how an insurer's staff can send data to generative AI tools.

  Stage

  Filed December 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A system implemented via a back-end application computer server of an enterprise, comprising:

  1. (a) a data store containing enterprise data
  2. (a) the back-end application computer server, coupled to the data store, including: a computer processor
  3. a computer memory coupled to the computer processor and storing instructions that, when executed by the computer processor, cause the back-end application computer server to: receive a prompt at at least one of an image component and a text component
  4. determine, via execution of at least one of the image component and the text component, a personal identifiable information (PII) status for the prompt, wherein: execution of the text component includes accessing a previously created and trained machine learning model, the machine learning model trained with at least one of internal data and internet content
  5. return a PII response
  6. receive selection of a large language model (LLM) in a case the PII status is PII-free
  7. determine a large language model status
  8. transmit the prompt to the selected LLM based on the large language model status
  9. receive a large language model (LLM) output
  10. (b) a communication port coupled to the back-end application computer server to facilitate an exchange of data with a remote device to support interactive user interface displays that provide information about the LLM output. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260093847A1 · filed 12/5/2025
• State Farm awaiting examination published 4/2/2026

  Systems and methods for enhanced instruction in virtual reality interactions

  Describes a virtual reality training system that reads sensor data while a trainee interacts with a client avatar, then applies a trained model to generate a scripted coaching message shown to the trainee during the exercise.

  How it works

  Presents a virtual environment with a client avatar to a trainee's device, receives sensor data from the trainee and client devices during an interaction, feeds that data into a trained machine learning model that outputs an instructional message containing scripted text for the trainee to say, and displays it on the trainee's device.

  Where it fits

  Simulated coaching for agents or claims staff who deal with customers.

  Stage

  Filed September 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A virtual reality computing system for conducting instructional interactions between one or more user computing devices including a trainee computing device within a virtual environment, the computing system comprising at least one memory device and at least one processor in communication with the at least one memory device, the at least one processor configured to:

  1. communicate with the one or more user computing devices including the trainee computing device associated with a trainee to cause the one or more user computing devices to present the virtual environment, wherein the virtual environment includes a client avatar representing a client interacting with the trainee in an instructional exercise
  2. receive sensor data from the trainee computing device associated with the trainee and a user computing device associated with the client during a current interaction between the trainee and the client within the virtual environment
  3. evaluate the current interaction between the trainee and the client by inputting the received sensor data into a trained machine learning (ML) model to generate one or more outputs including an instructional message that includes scripted text for the trainee to communicate to the client during the current interaction within the virtual environment
  4. present, on the trainee computing device, the instructional message. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260093375A1 · filed 9/16/2025
• EXL awaiting examination published 3/26/2026

  Platform for artifact generation using a trained neural network

  Describes a platform that turns transcripts and audiovisual recordings into tokens and prompts a trained neural network to produce a requested artifact such as a call summary or requirements document, with a chatbot to query the result.

  How it works

  Processes a set of video frames into tokens, captures an output-type instruction naming the artifact wanted, has a trained neural network generate that artifact conditioned on project type, user role, domain or technology stack, and displays it beside a chatbot whose context is set to the artifact so a user query searches it.

  Where it fits

  General document and meeting-artifact generation; no specific insurance function is described.

  Stage

  Filed September 2024 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer-implemented method for automatically identifying content used to generate project artifacts from transcripts and audiovisual files, the computer-implemented method comprising:

  1. processing a set of video frames to generate a set of tokens for a project artifact
  2. using a graphical user interface (GUI), capturing an output type instruction indicative of a specific output type of the project artifact, wherein the specific output type includes one or more of: a call summary, meeting minutes, a user story, a test case, a business requirements document, or process steps
  3. using the output type instruction and the generated set of tokens, causing a trained neural network to generate the project artifact according to: (i) the specific output type and (ii) at least two of: a project type, a user role, a domain, or a technology stack descriptor, wherein the project artifact comprises a body of text
  4. causing the GUI to display: (i) a first component comprising the generated project artifact and (ii) a second component comprising a chat bot having a context for the chat bot set to the generated project artifact
  5. responsive to detecting a user query at the GUI, causing the chat bot to search the generated project artifact displayed in the first component using the detected user query to generate a set of search results
  6. displaying the generated set of search results at the second component of the GUI. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260087229A1 · filed 9/26/2024
• State Farm awaiting examination published 3/19/2026

  Policyholder setup in secure personal and financial information storage and chatbot access by trusted individuals

  Describes a system for storing a policyholder's financial and account information and later guiding a trusted individual through the estate using a chatbot that simulates a characteristic of the account owner.

  How it works

  Receives information identifying a user authorized to access a digital asset, instructs an electronic device to connect to a digital asset service, receives a request to use the service, establishes a virtual communication session with the device, and during the session provides a simulation of a characteristic of the account's owner.

  Where it fits

  Estate and beneficiary servicing tied to life or financial policies.

  Stage

  Filed November 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer-implemented method for configuring a computing system to provide at least one of digital asset storage, management, or retrieval services associated with a digital account, the computer-implemented method comprising:

  1. receiving, by a processor, information identifying a user authorized to access a digital asset associated with the digital account
  2. sending, by the processor, based on the information, and to an electronic device, an instruction causing the electronic device to connect to a digital asset service
  3. receiving, by the processor and based on the electronic device connecting to the digital asset service, a request to use the digital asset service
  4. establishing, by the processor and based on the request, a virtual communication session with the electronic device via a computer program
  5. providing, by the processor and during the virtual communication session, a simulation of a characteristic of an owner of the digital account. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260080489A1 · filed 11/17/2025
• State Farm awaiting examination published 3/19/2026

  Apparatuses, systems, and methods for determining vehicle operator distractions at particular geographic locations

  Describes a device that reads in-cabin sensor data to gauge occupant distraction from torso posture and its weighted duration, then computes a probability of distraction at a particular geographic location for driving-risk assessment.

  How it works

  One module receives interior sensor data, where distraction severity scales with the degree of motion of the occupant's torso posture and its weighted duration, and a second module combines that data with vehicle location data to generate distraction data representing the probability of occupant distraction at a particular location.

  Where it fits

  Telematics-based driving risk scoring for usage-based auto pricing.

  Stage

  Filed November 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A device for determining vehicle occupant distractions at particular geographic locations, the device comprising:

  1. a vehicle interior data receiving module stored on a memory that, when executed by a processor, causes the processor to receive vehicle interior data from at least one vehicle interior sensor, wherein the vehicle interior data is representative of at least one distraction of at least one vehicle occupant, and wherein a severity of a distraction is proportional to a degree of motion of a respective vehicle occupant torso posture and a weighted duration time of the respective vehicle occupant posture, the vehicle occupant posture including at least vehicle occupant torso and joint data points associated with the at least vehicle occupant torso
  2. a vehicle occupant distraction data generation module stored on the memory that, when executed by the processor, causes the processor to generate vehicle occupant distraction data based on the vehicle interior data and vehicle location data, wherein the vehicle occupant distraction data is representative of a probability of at least one vehicle occupant distraction at a particular geographic location, and wherein the probability is proportional to the severity of the distraction according to the degree of motion of the respective vehicle occupant posture and the weighted duration time of the respective vehicle occupant posture. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260077771A1 · filed 11/19/2025
• State Farm awaiting examination published 3/12/2026

  Representative client devices in a contact center environment

  Describes a contact-center framework in which a server aggregates interaction data across representatives' sessions and returns operating instructions that drive a representative's device to start a call or take up a deferred task.

  How it works

  A server determines interaction data from actions in two separate representative-customer sessions, generates operating instructions from that combined data, and transmits them to a representative device, causing it to either start an additional session through an external provider or begin a deferred work item from an internal provider.

  Where it fits

  Contact-center operations for insurance customer service.

  Stage

  Filed November 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A method comprising:

  1. determining, by a server, first interaction data associated with a first action performed during a first communication session between a first representative device and a first customer device
  2. determining, by the server, second interaction data associated with a second action performed during a second communication session between a second representative device and a second customer device
  3. generating, by the server and based on the first interaction data and the second interaction data, operating instructions executable by a processor of the first representative device
  4. transmitting, by the server, the operating instructions to the first representative device, the executable operating instructions causing the processor of the first representative device to perform at least one of: initiating, on the first representative device, an additional communication session via an external communication service provider
  5. or initiating, on the first representative device, a deferred work item received from an internal service provider. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260075144A1 · filed 11/17/2025
• CVS Health awaiting examination published 3/12/2026

  Systems and methods for federated knowledge goverance

  Describes a federated knowledge-base platform that resolves a search request to the certified version of a document held in one of several underlying systems, using a stored snapshot reference to locate and retrieve it.

  How it works

  A consumer device requests a certified version of a document, the federated platform matches it to a snapshot reference identifying the storage location within a particular knowledge-base system, retrieves that identifier, uses it to obtain the certified version from that system, and builds and displays an interface showing the document.

  Where it fits

  Document governance across systems; no specific insurance function is described.

  Stage

  Filed September 2024 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A system, comprising:

  1. a consumer device that is configured to provide a search request for a certified version of a first document that is stored in a first knowledge base (KB) computing system
  2. a federated KB computing platform configured to: receive the search request from the consumer device
  3. based on the search request, determine a first document snapshot reference that comprises an identifier that identifies a storage location of the certified version of the first document within the first KB computing system, wherein the federated KB computing platform stores a plurality of document snapshot references associated with a plurality of documents that are stored in a plurality of different KB computing systems, and wherein the first KB computing system stores a plurality of versions of the first document including a latest version of the first document and the certified version of the first document
  4. based on the determination, retrieve the identifier that identifies the storage location of the certified version of the first document within the first KB computing system
  5. based on providing the identifier to the first KB computing system, obtain the certified version of the first document from the first KB computing system
  6. generate a user interface (UI) based on the retrieved certified version of the first document
  7. cause display of the UI on the consumer device, wherein the consumer device is configured to display the UI comprising the retrieved certified version of the first document
  8. the first KB computing system configured to: receive the identifier from the federated KB computing platform
  9. retrieve the certified version of the first document using the identifier
  10. provide the certified version of the first document to the federated KB computing platform. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260073244A1 · filed 9/6/2024
• Travelers awaiting examination published 2/26/2026

  Systems and methods for mixed reality (mr) and artificial intelligence (AI)-enhanced spatial investigation

  Describes a mixed-reality headset with depth and motion sensing that builds a 3-D model of a fire-damaged site as the wearer moves through it and runs AI spatial and safety models over the scene for claim investigation.

  How it works

  A head-mounted see-through display with a time-of-flight sensor, camera and inertial units captures distances to surface points from one location, computes a point-cloud portion, tracks the wearer's movement to a second location, captures and computes a second portion, and merges them into a 3-D wire-mesh model of the environment for analysis.

  Where it fits

  On-site property and fire loss investigation for claims.

  Stage

  Filed October 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A system for Mixed Reality (MR) and Artificial Intelligence (AI)-enhanced spatial investigation, comprising:

  1. an MR device in wireless communication with the sensor device, the MR device, comprising: a processing device
  2. a wireless communication device in communication with the processing device and in selective communication with the sensor device
  3. a head-mounted see-through display in communication with the processing device
  4. a Time of Flight (ToF) sensor in communication with the processing device
  5. a camera in communication with the processing device
  6. one or more Inertial Measurement Unit (IMU) devices in communication with the processing device
  7. a battery in communication with the processing device
  8. a memory device in communication with the processing device, the memory device storing MR instructions that when executed by the processing device, result in: acquiring, by the ToFsensor and at a first time and from a first location in an environment in which a wearer of the head-mounted see-through display is located, data descriptive of first distances from the MR device to a first plurality of surface points in the environment, wherein the first plurality of surface points are within a first field of view of the camera
  9. computing, by the processing device and utilizing the first distances and the first location, a first portion of a 3-D point cloud descriptive of locations of the first plurality of surface points in the environment
  10. tracking, after the acquiring of the first distances and by the one or moreIMU devices, a first movement of the wearer from the first location in the environment to a second location in the environment
  11. acquiring, by the ToFsensor and at a second time and from the second location in the environment, data descriptive of second distances from the MR device to a second plurality of surface points in the environment, wherein the second plurality of surface points are within a second field of view of the camera
  12. computing, by the processing device and utilizing the second distances and the second location, a second portion of the 3-D point cloud descriptive of locations of the second plurality of surface points in the environment
  13. generating, by the processing device and utilizing the first and second portions of the 3-Dpoint cloud, a 3-D wire mesh model descriptive of the environment
  14. receiving, from the se

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260057674A1 · filed 10/23/2025
• Centene under examination published 2/19/2026

  Artificial intelligence-based personalized care delivery systems and methods

  Describes a health-plan system that applies a member's health data to an observation model to identify conditions, feeds those to service models that generate a treatment plan, and notifies both a provider and the member of a recommended appointment.

  How it works

  Receives member health data, applies it to an observation model trained to identify the member's conditions from historical data, passes those conditions to service models trained to generate treatment plans, takes a recommended plan with an appointment, and sends appointment notifications to both a care-provider device and the member's device.

  Where it fits

  Care management and utilization steering for a health plan.

  Stage

  Filed August 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A personalized health plan (PHP) computing system comprising at least one processor in communication with at least one memory, wherein the at least one processor is configured to:

  1. receive member health data associated with a member from at least one of a database or a member computing device
  2. apply the member health data to an observation model, wherein the observation model is trained to identify health conditions of the member based upon historical health data
  3. receive a health conditions output from the observation model, the health conditions output including one or more health conditions associated with the member based upon the member health data
  4. apply the health conditions output to one or more service models, wherein the one or more service models are trained to generate treatment plans associated with care for different health conditions
  5. receive a service response output from the service model, the service response output including a recommended treatment plan comprising a recommended health appointment for the member based upon a health condition of the one or more health conditions
  6. cause a first notification for the recommended health appointment to be transmitted to a care provider computing device associated with a care provider to conduct the recommended health appointment for the member
  7. cause a second notification for the recommended health appointment to be transmitted to the member computing device to notify the member of the recommended health appointment with the care provider. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 9/3/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  References cited against it: Odessky et al, Kohsla et al, Mason et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260051405A1 · filed 8/18/2025
• Nearmap awaiting examination published 2/19/2026

  Artificial intellegence-based property information processing system and method pinhole training

  Describes a two-network method for generating per-pixel semantic predictions from imagery: a first CNN trained on a sparse grid of pixels, from which a second, modified CNN is derived to predict labels at higher pixel density.

  How it works

  Defines a first CNN over a mask of pixels spaced s apart and trains it on image crops against target labels by minimizing prediction error, then builds a second CNN from the first's weights with a reduced final-layer stride and dilated later layers, and runs it on a full image to output semantic predictions at higher density than the mask.

  Where it fits

  Aerial-imagery feature extraction feeding property risk assessment; the claim itself is a general image-segmentation technique.

  Stage

  Filed August 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A method of generating semantic predictions corresponding to pixels of an input image, executing on a computer having a network connection, the method comprising the steps of:

  A method of generating semantic predictions corresponding to pixels of an input image, executing on a computer having a network connection, the method comprising the steps of: providing a first Convolutional Neural Network (CNN) by: selecting a sampling factor, s, defining a mask that corresponds to a grid of pixels positioned s pixels apart from each other and centered on a single pixel of the grid of pixels, defining an input image crop size based on the extent of the mask, defining a first CNN that processes image data shaped according to the input image crop to generate semantic predictions for pixels corresponding to the grid of pixel locations of the mask, receiving a training image comprising image pixels, generating a training image crop according to the input image crop size, receiving target semantic labels for the training image, generating a grid of target semantic labels for the training image crop corresponding to the grid of pixels of the mask, processing, using the first CNN, the training image crop to generate a grid of output semantic predictions for the training image crop corresponding to the mask pixels of the training image crop, determining a difference between the target semantic labels and the output semantic predictions, and adjusting parameters of the first CNN to reduce a difference between the target semantic labels and the output semantic predictions, providing a second CNN according to weight tensors and an architecture of the first CNN, the second CNN including modified convolution layers, wherein a last convolution layer has a stride that is reduced by dividing a scaling factorf, and all convolution layers after the last convolution layer are dilated by multiplying the scaling factor f , receiving with the second CNN the input image that comprises pixels, and generating with the second CNN semantic predictions for the input image that is at a higher pixel density than the mask. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260051022A1 · filed 8/12/2025
• Nearmap awaiting examination published 2/19/2026

  Roof age determination using AI models

  Describes an AI system that turns roof imagery and non-image data into vector representations and compares them against prior vectors to detect whether a roof has changed or to estimate its age, supporting property risk assessment.

  How it works

  Receives geospatial data tiles each carrying a temporal identifier and location, processes a subset to derive attribute data, generates attribute keys from the identifier and processing configuration, stores the attributes by key, and on a property request selects the overlapping tile, computes its key and returns its stored attributes.

  Where it fits

  Roof condition and age inputs for property underwriting and renewals.

  Stage

  Filed August 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A method of generating, storing and accessing attribute data corresponding to geospatial data tiles, the method executing on a computer having a network connection and having software executing thereon to perform the steps of:

  1. receiving a set of geospatial data tiles via the network connection, the geospatial data tiles saved on a storage accessible by the computer, each geospatial data tile having a temporal identifier and location data
  2. processing a subset of the geospatial data tiles to determine attribute data corresponding to features in the geospatial data tiles according to a processing configuration
  3. generating attribute data keys based on the temporal identifier of each of the geospatial data tiles and the processing configuration
  4. storing the attribute data for each geospatial data tile on the storage at a location defined, at least based in part, on the attribute data key
  5. receiving location information identifying a property and a request for attribute data corresponding to the property
  6. determining a geographical region corresponding to the property
  7. selecting a geospatial data tile from the set of geospatial data tiles for which the location data overlaps with the geographical region
  8. associating attribute data of the request with a processing configuration of the request
  9. determining an attribute data key for the selected geospatial data tile based on its temporal identifier and the processing configuration of the request
  10. SVG 19297154.08-12-2025.ME8USJB7X224X60.CLM.1.22.2047.2724.2061.2747.svg 0.077 0.047 Chemistry Black and white identifying previously computed attribute data stored on the storage at a location based Page 1 of 5 on, at least in part, the attribute data key for the selected geospatial data tile
  11. transmitting data to the user computer, the transmitted data based on the identified attribute data. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260050996A1 · filed 8/12/2025
• State Farm under examination published 2/12/2026

  Residential building remaining useful life detector

  Describes a system that estimates the remaining lifespan of a home's plumbing from residential records via a trained model, then revises the estimate after the homeowner reports taking recommended actions.

  How it works

  Receives plumbing and property aspects from residential records, has a trained model estimate each aspect's impact to predict the plumbing's remaining lifespan, issues a notification with that lifespan and recommended actions, then on user input that actions were taken computes a lifespan modification factor and issues an updated estimate.

  Where it fits

  Property risk assessment and loss prevention for homeowners cover.

  Stage

  Filed October 2024 and under examination; the claim read here was already narrowed by amendment, not the original.

  The claim as amended, showing changes

  A computer-implemented method for evaluating aspects of a plumbing system of a residential property using at least one processor in communication with at least one memory device, the computer-implemented method comprising:

  1. communicating the address to an external database, the external database containing a plurality of residential records associated with the address in a geographic location
  2. receiving, via a user device, a first user input indicative of an address for the residential property
  3. receiving, via a _ [[the ]]plurality of residential records, a plurality of plumbing system aspects and a plurality of residential property aspects
  4. predicting, via a trained machine learning model configured to estimate a respective impact of the plurality of plumbing system aspects and the plurality of residential property aspects on a remaining lifespan of the plumbing system, an estimated remaining lifespan of the plumbing system based upon the plurality of plumbing system aspects and the plurality of residential property aspects
  5. generating a first notification, the first notification comprising the estimated remaining lifespan of the plumbing system and one or more recommended actions
  6. receiving a seconduser input indicating a _ [[the ]]user has implemented the one or more recommended actions
  7. generating a lifespan modification factor based upon the one or more recommended actions, wherein the lifespan modification factor is indicative of an effect the one or more recommended actions has on the estimated remaining lifespan of the plumbing system
  8. and generating and transmitting, to a _ [[the ]]user device of the user for display, a second notification comprising an updated estimated remaining lifespan of the plumbing system and one or more additional recommended actions, the updated estimated remaining lifespan beingbased upon the lifespan modification factor and the estimated remaininglifespan indicative of an effect the one or more recommended actions has on the estimated remaining lifespan of the plumbing system. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 3 office actions

  First action 8/26/2025, most recent 6/5/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Parris et al, Fang et al, Parris, Fang.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260044835A1 · filed 10/17/2024
• State Farm awaiting examination published 2/12/2026

  Residential building remaining useful life detector

  Describes a system that processes several types of residential data through a trained model to produce a plumbing impact score representing the remaining useful life of a home's plumbing, and triggers a follow-up action.

  How it works

  Receives several different types of residential data about a plumbing system, has a trained machine learning model estimate each type's impact on remaining useful life and generate a residential plumbing impact score predicting that life, and initiates an action in response to the score.

  Where it fits

  Property risk assessment and maintenance prompts for homeowners cover.

  Stage

  Filed October 2024 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer system for assessing a remaining useful life of a residential plumbing system of a residential building, the residential plumbing system comprising one or more plumbing components of the residential building, the system comprising:

  1. one or more memory devices having instructions stored thereon that, when executed by one or more processors, cause the one or more processors to perform operations comprising: receiving a plurality of different types of residential data associated with the residential plumbing system
  2. determining, by processing the plurality of different types of residential data using a trained machine learning model, a residential plumbing impact score, the residential plumbing impact score indicating the remaining useful life of the residential plumbing system, wherein the trained machine learning model is configured to: estimate an impact of the plurality of different types of residential data on the remaining useful life of the residential plumbing system
  3. generate the residential plumbing impact score by predicting the remaining useful life of the residential plumbing system using the estimated impacts of the plurality of different types of residential data on the remaining useful life
  4. initiating an action relating to the residential plumbing system responsive to the generation of the residential plumbing impact score. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260044778A1 · filed 10/17/2024
• State Farm awaiting examination published 2/5/2026

  Systems and methods for property damage prevention and mitigation

  Describes a system that trains and retrains models on property parameters and telematics data to estimate a property's damage likelihood, identify the factors driving it and recommend mitigating actions.

  How it works

  Trains a model on property parameters, receives telematics data from sensors on property devices, retrains the model on that data, outputs a likelihood of damage from how the devices are functioning, determines the damage factors that raise that likelihood, and outputs recommended actions to mitigate it.

  Where it fits

  Loss prevention and property risk assessment for homeowners cover.

  Stage

  Filed October 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computing system for training computer models associated with functionality of property devices, the computing system comprising at least one processor in communication with a memory device and one or more telematics sensors associated with one or more property devices of a property, the at least one processor configured to:

  1. train, using machine learning tools or artificial intelligence, one or more computer models by inputting a plurality of property parameters into the one or more computer models
  2. receive, via the one or more telematics sensors, property telematics data associated with a functioning of the one or more property devices
  3. re-train, using the machine learning tools or artificial intelligence, the one or more trained computer models by inputting the property telematics data into the one or more computer models
  4. output, from the one or more re-trained models, a likelihood of damages to the property due to the functioning of the one or more property devices
  5. determine, using the one or more re-trained computer models, one or more damage factors including at least one aspect of the property that increases the likelihood of the property incurring the damages
  6. output, from the one or more re-trained computer models, one or more recommended actions associated with the functioning of the one or more property devices for mitigating the likelihood of the property incurring the damages. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260038056A1 · filed 10/10/2025
• State Farm allowed published 2/5/2026

  Approving and updating dynamic mortgage applications

  Describes a method for approving a loan tied to a property by retrieving customer information from two blockchains, verifying it cryptographically, determining an approval status and recording that status on-chain.

  How it works

  Receives a loan request tied to a property, retrieves first customer information from a block of one blockchain and second information from another, verifies the second via a cryptographic key-pair by reaching consensus with a remote node, determines an approval status, writes it into a new block and verifies it with a second key-pair.

  Where it fits

  Mortgage approval workflows; the claim centers on blockchain verification rather than an insurance function.

  Stage

  Filed August 2025; the application is at the allowed stage on the original claims as filed.

  The claim as filed

  A computer-implemented method for determining a customer is approved for a loan, the computer-implemented method comprising:

  1. receiving, at a processor, a request for a loan associated with a real estate property
  2. retrieving, by the processor, based at least in part on information associated with the request, and from a first block of a first blockchain, first customer information
  3. retrieving, by the processor and from a second block of a second blockchain, second customer information
  4. verifying, by the processor, the second customer information based at least in part on a first cryptographic key-pair associated with the second blockchain by establishing a consensus with a remote node that a solution associated with the first cryptographic key-pair is valid
  5. based at least in part on verifying the second customer information, determining, by the processor and based at least in part on the second customer information and the request for the loan, an approval status for the loan
  6. generating, by the processor and based at least in part on the approval status, a third block of the first blockchain comprising data representing the approval status
  7. verifying, by the processor, the approval status based at least in part on a second cryptographic key-pair associated with the first blockchain. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260038037A1 · filed 8/7/2025
• MassMutual awaiting examination published 2/5/2026

  Systems and methods for processing electronic requests

  Describes a server that validates the inputs to a predictive model against a specification file, imputes valid values when the inputs do not conform, and then executes the model to return results.

  How it works

  Receives a request to run a predictive model with a set of inputs, validates the inputs against required ranges or types in the model's specification file, and when they do not conform computes valid values, revises the model's package to include them and an identifier, executes the model within that revised package and returns the output.

  Where it fits

  Operational plumbing for running actuarial or predictive models on request; no specific insurance line is described.

  Stage

  Filed August 2025 and awaiting examination; the original claims were cancelled and this was submitted as a new claim in their place.

  The claim new claim

  A method comprising:

  1. receiving, by at least one processor, from an electronic device, an electronic request to execute a first predictive computer model of a plurality of predictive computer models using a set of inputs, the first predictive computer model having a package comprising at least one required input range or type based on a specification file comprising a set of validation codes of the first predictive computer model
  2. validating, by the at least one processor, prior to executing the first predictive computer model, the electronic request by verifying that the set of inputs corresponds to the at least one required input range or type
  3. in response to the set of inputs not corresponding to the at least one required input range or type in the set of validation codes of the first predictive computer model: calculating, by the at least one processor, one or more valid values for the set of inputs within the at least one required input range or type
  4. revising, by the at least one processor, the package of the first predictive computer model to include the one or more valid values and an identifier of the package
  5. executing, by at least one processor, the first predictive computer model based on the one or more valid values by calling functions of the first predictive computer model within its respective revised package to generate output results using the identifier
  6. transmitting, by the at least one processor, the output results to at least one electronic device. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260037838A1 · filed 8/12/2025
• Zesty.ai awaiting examination published 2/5/2026

  Wind Predictions Using Artificial Intelligence

  Describes a method that chains machine learning models to estimate wind speed and report frequency for a location and then derive damage frequency and severity metrics, supporting catastrophe risk assessment.

  How it works

  Receives a location, has a wind-speed model and a wind-report-frequency model estimate those values, feeds the values plus location feature data into a damage-frequency model to produce a damage-frequency metric, and feeds a separate feature set into a damage-severity model to produce a damage-severity metric.

  Where it fits

  Wind and catastrophe risk assessment for property pricing and underwriting.

  Stage

  Filed July 2024 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer implemented method comprising:

  1. receiving, using one or more processors, a location
  2. determining, using the one or more processors, a wind speed associated with the location using a first wind speed machine learning model
  3. determining, using the one or more processors, a wind report frequency associated with the location using a first wind report frequency machine learning model
  4. obtaining, using the one or more processors, first feature data associated with the location, the first feature data including the wind speed associated with the location, the wind report frequency associated with the location, and data describing a first set of features at the location
  5. determining, using the one or more processors, a damage frequency metric associated with the location by applying a first damage frequency machine learning model to the first feature data
  6. obtaining, using the one or more processors, second feature data associated with the location, the second feature data including data describing a second set of features at the location
  7. determining, using the one or more processors, a damage severity metric associated with the location by applying a first damage severity machine learning model to the second feature data. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260036715A1 · filed 7/30/2024
• Elevance under examination published 1/29/2026

  Systems and Methods for Predicting Outcomes Using Large Language Models

  Describes training a causal language model on tokenized, chronologically ordered clinical event codes to predict outcomes, and generating a synthetic dataset for evaluation.

  How it works

  Obtains a training set of coded data, preprocesses it by chronologically ordering each individual's event codes, inserting delimiter tokens to separate intra-event, inter-event and per-individual data, and tokenizing with a fixed vocabulary, then trains a causal language model to predict the next code as a probability distribution.

  Where it fits

  Outcome prediction from medical claim and clinical codes for a health plan.

  Stage

  Filed July 2025 and under examination; these are the original claims as filed.

  The claim as filed

  A method of training a causal language model for predicting outcomes, the method comprising:

  1. obtaining a training dataset that includes structured data including codes
  2. preprocessing the structured data to convert raw event requests into a structured token sequence, including: performing a sorting algorithm to organize codes within each event request in the structured data into a clinically logical sequence, including chronologically ordering event requests for a respective individual to form a temporally sequenced dataset thereby enabling a machine learning model to learn chronological order of events
  3. inserting one or more delimiter tokens into the structured data for concatenating intra-event request codes, inter-event request codes for the respective individual, and data for different individuals, thereby enabling batch data processing
  4. tokenizing the structured data using a tokenizer to obtain a sequence of tokens, wherein the tokenizer preserves the one or more delimiter tokens to maintain context of event request data, wherein the tokenizer is trained on event request data with a predetermined vocabulary size
  5. training a causal language model using the structured token sequence to predict an outcome, wherein training the causal language model comprises predicting a next code in the structured token sequence based on prior codes, thereby generating a sequence of codes for each event request in a causally coherent manner, wherein predicting the next code is modeled as a probability distribution over possible codes. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/27/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Rose et al.

  A further rejection under section 101 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260031234A1 · filed 7/28/2025
• State Farm awaiting examination published 1/29/2026

  Method of controlling for undesired factors in machine learning models

  Describes a method that trains a model and then retrains it to identify and exclude undesired factors such as age, sex or race from its output when analyzing an applicant's image or audio during underwriting.

  How it works

  Trains a first model whose output includes one or more undesired factors, identifies those factors in the output, retrains the model to identify and exclude them to produce a second model, and runs the second model to analyze an individual's images or audio while excluding those factors from its output.

  Where it fits

  Underwriting from image or audio while suppressing protected-class factors.

  Stage

  Filed October 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer system configured to train a machine learning model, the computer system comprising at least one processor configured to:

  1. train the machine learning model using a training data set to produce a first trained machine learning model that generates a first output that includes one or more undesired factors
  2. identify the one or more undesired factors included in the first output of the first trained machine learning model
  3. train the first trained machine learning model based upon the identified one or more undesired factors to produce a second trained machine learning model trained to identify and exclude the identified one or more undesired factors from a second output
  4. run the second trained machine learning model to analyze at least one of images or audio of an individual while excluding some or all of the identified one or more undesired factors from the second output. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260030685A1 · filed 10/3/2025
• Hartford final rejection published 1/29/2026

  Framework for query generation in an artificial intelligence environment

  Describes a framework that turns a natural language query into a SQL query using a large language model trained on an application-specific data dictionary, runs it through an API, and returns a natural language response.

  How it works

  Builds a per-application data dictionary of each object's name, type, size and relationships, uses a large language model trained on it to turn a natural language query into a SQL query, calls a no-contract API to run it, and depending on whether the response answers the query returns a natural language answer or generates a second SQL query.

  Where it fits

  Natural language access to insurer data stores; no specific line of business is described.

  Stage

  Filed July 2024 and standing at a final rejection; the claim read here was already narrowed by amendment, not the original.

  The claim as amended, showing changes

  A system comprising:

  1. a memory storing program code: and one or more processing units to execute the program code to cause the system to: generate an application-specific data dictionary for each application, the application-specific data dictionary including for each of a plurality of objects: an object name, a description, a data type, a size, a classification and a relationship with other data sets, wherein the description includes terms used to describe the object and generate at least a first Structured Query Lan guag e (SQL) query
  2. receive a natural language query
  3. generate a Structured Query Language (SQL)a first SQL query based on the received natural language query and using a large language model (LLM) trained with [[an]] the application-specific data dictionary, the generated SQL query including an endpoint and one or more fields based on the application- specific data dictionary trained LLM
  4. determine, via a no contract-based Application Programming Interface (API), the endpoint and an API call from data included in the generated SQL query
  5. invoke the no contract-based API
  6. receive a response to the firstSQL query from a data source via the no contract-based API
  7. determine the natural language query is answerable with the received response
  8. generate a second SQLquery in a case it is determined the natural lan guage q uery is determined unanswerable, wherein the second SQLquery is one o f: a newquery and an update to the first SQLquery
  9. generate a natural language response from the response to the SQL query in a case the natural language query is determined answerable
  10. transmit the natural language response to an entity. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 4 office actions

  First action 7/9/2025, most recent 7/23/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Zhao et al, Hoang et al, US 12,321,791, Guan et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260030299A1 · filed 7/26/2024
• State Farm awaiting examination published 1/22/2026

  Insurance claim processing in secure personal and financial information storage and chatbot access by trusted individuals

  Describes a system that opens a conversation session with a beneficiary, verifies the customer's identity from a certifying digital file, and processes a claim-related digital file as part of estate handling.

  How it works

  Receives a digital file certifying an event and a request to access a customer's asset, establishes a conversation session with the beneficiary, verifies that the identity in the request matches the customer's actual identity, and on a match generates a second file and returns information indicating the action needed to process it.

  Where it fits

  Claim intake and estate handling for beneficiaries on a policy.

  Stage

  Filed September 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer-implemented method for establishing a conversation session with a beneficiary and processing a digital file, the computer-implemented method comprising:

  1. receiving, by a processor and from a computing device, a first digital file and a request to access an asset of a customer, the request indicating an identity of the customer and an occurrence of an event associated with the customer, and the first digital file certifying the event
  2. based on receiving the request and the first digital file, establishing, by the processor, a conversation session with a beneficiary of the customer via the computing device
  3. verifying, by the processor and via the conversation session, the identity indicated in the request matches an actual identity of the customer
  4. based on verifying the identity matches the actual identity, generating, by the processor, a second digital file corresponding to the request
  5. transmitting, by the processor and to the computing device, information indicating an action required for processing the second digital file, wherein receipt of the information by the computing device causes a display of the computing device to indicate the action. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260024154A1 · filed 9/25/2025
• State Farm awaiting examination published 1/22/2026

  Reducing false positives using customer feedback and machine learning

  Describes a method that generates fraud-classification rules with a trained model and applies them to account data to identify the type of fraud, aimed at reducing false positive alerts.

  How it works

  A trained machine learning program, trained on data indicating types of fraud associated with transactions or accounts, generates fraud-classification rules, and the system accesses a particular account's data and applies those rules to output a classification naming the type of fraud associated with that account.

  Where it fits

  Fraud detection and alert tuning in claims or financial servicing.

  Stage

  Filed September 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A computer-implemented method for classifying a type of fraud associated with an account based on account data and fraud classification rules, comprising:

  1. generating, by a computing system comprising a processor, the fraud classification rules, wherein: the fraud classification rules are generated using a trained machine learning program trained based on a data set indicating types of fraud, of a set of different types of fraud, associated with transactions or accounts
  2. accessing, by the computing system, the account data associated with a particular account
  3. generating, by the computing system, and by applying the fraud classification rules to the account data, a fraud classification that identifies a particular type of fraud, included in the set of different types of fraud, associated with the particular account. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260024101A1 · filed 9/30/2025
• LexisNexis Risk awaiting examination published 1/22/2026

  Systems and methods for automatic identification of text fields and data type

  Describes a method that classifies a webpage's text fields into functional groups from their metadata and behavior, then selects a matching behavioral analysis algorithm to flag authentication issues or fraud.

  How it works

  Monitors a webpage's text fields, receives their field data and metadata, builds feature vectors of the fields' contextual patterns, classifies the fields into functional groups, selects a behavioral analysis algorithm from those groups, processes the field data with it, and outputs an indication of user authentication or potential fraud.

  Where it fits

  Identity verification and fraud screening in online journeys.

  Stage

  Filed September 2025 and awaiting examination; these are the original claims as filed.

  The claim as filed

  A method for processing text field data associated with online website interactions, the method comprising:

  1. monitoring one or more text fields of a webpage
  2. receiving field data and metadata associated with the one or more text fields
  3. constructing, from the field data and metadata, one or more feature vectors representing contextual patterns of the text fields
  4. classifying the field data into functional groups based on the one or more feature vectors
  5. automatically selecting a behavioral analysis algorithm based on the classified functional groups
  6. processing the field data using the selected behavioral analysis algorithm
  7. outputting an indication of user authentication or potential fraudulent activity based on the processing. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260023830A1 · filed 9/25/2025
• State Farm under examination published 1/22/2026

  System and method for automatically monitoring and diagnosing user experience problems

  Describes a method that extracts a website's click data, processes it into weighted n-grams, and uses a trained analytics engine to diagnose which user-experience problem from a candidate list the site has.

  How it works

  Receives a list of potential user-experience problems, extracts click data from a website, processes it into weighted grams whose weights come from click timing or webpage frequency or into forward and reverse grams, and uses a trained analytics engine to diagnose which listed problem the site has.

  Where it fits

  Diagnosing digital self-service issues on an insurer's website.

  Stage

  Filed July 2025 and under examination; these are the original claims as filed.

  The claim as filed

  A computer-implemented method for diagnosing a problem of a website, the website comprising a plurality of webpages, and the method comprising, via one or more processors:

  1. receiving a list of potential user experience problems
  2. extracting click data from the website
  3. receiving processed click data, the click data having been processed by processing the extracted click data into: (i) weighted grams, the weighted grams comprising a weight: (a) assigned to a gram as a whole based on a time derived from the click data, or (b) based on a webpage frequency derived from the click data
  4. or (ii) at least one forward gram, and at least one reverse gram
  5. using a trained analytics engine to diagnose the problem of the website with a potential user experience problem of the received list of potential user experience problems. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/10/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260023639A1 · filed 7/31/2025
• Travelers awaiting examination published 1/8/2026

  Systems and methods for mixed reality (mr) and artificial intelligence (AI)-enhanced fire investigation

  Describes a mixed reality and AI fire investigation system that builds a 3-D model of a fire-damaged location, runs AI fire and safety analyses, and layers data onto the model, supporting claims investigation.

  How it works

  A time-of-flight sensor on a head-mounted see-through display captures distances to surface points as the wearer moves, an inertial measurement unit tracks that movement between locations, and the device stitches successive point-cloud portions into a 3-D wire mesh model of the fire-damaged environment.

  Where it fits

  Property claims investigation and loss adjustment for fire-damaged structures.

  Stage

  Filed September 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A method for Mixed Reality (MR) and Artificial Intelligence (AI)-enhanced fire investigation, comprising:

  1. acquiring, by a Time of Flight (ToF) sensor of an MR device, the ToFsensor being in communication with a processing device of the MR device, and at a first time and from a first location in an environment in which a wearer of a head-mounted see-through display of the MR device is located, the head-mounted see-through display being in communication with the processing device, data descriptive of first distances from the MR device to a first plurality of surface points in the environment, wherein the first plurality of surface points are within a first field of view of a camera of the MR device, the camera being in communication with the processing device
  2. computing, by the processing device and utilizing the first distances and the first location, a first portion of a 3-D point cloud descriptive of locations of the first plurality of surface points in the environment
  3. tracking, after the acquiring of the first distances and by one or more Inertial Measurement Unit (IMU) devices of the MR device, the one or moreIMU devices being in communication with the processing device, a first movement of the wearer from the first location in the environment to a second location in the environment
  4. acquiring, by the ToFsensor and at a second time and from the second location in the environment, data descriptive of second distances from the MR device to a second plurality of surface points in the environment, wherein the second plurality of surface points are within a second field of view of the camera
  5. computing, by the processing device and utilizing the second distances and the second location, a second portion of the 3-D point cloud descriptive of locations of the second plurality of surface points in the environment
  6. generating, by the processing device and utilizing the first and second portions of the 3-D point cloud, a 3-D wire mesh model descriptive of the environment
  7. receiving, by a wireless communication device of the MR device, the wireless communication device being in communication with the processing device, and from a sensor device in selective communication with the wireless communication device, (i) data descriptive of the environment that has been captured by the sensor device and (ii) positioning information descriptive of a location and orientatio

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260011151A1 · filed 9/11/2025
• TIAA awaiting examination published 1/8/2026

  Optimizing enterprise technology ecosystems using artificial intelligence to analyze data, predict stability, identify dependencies, and generate actionable tasks

  Describes a system that collects enterprise data, applies AI to analyze it, predicts stability, identifies dependencies, and generates actionable tasks to maintain equilibrium across a technology ecosystem.

  How it works

  It collects logging and monitoring data across the ecosystem's functional layers, scores each layer's stability with one machine learning model and rolls those into an overall health score with a second, then uses a generative AI model to synthesize tickets and epics as remediation tasks and refreshes its data sources on event triggers.

  Where it fits

  Enterprise technology operations and reliability rather than a specific insurance workflow.

  Stage

  Filed July 2024, awaiting examination, written from a claim the applicant has already amended during prosecution.

  The claim as amended, showing changes

  A computing system comprising:

  1. a memory having stored thereon computer-executable instructions that, when executed by the processor, cause the computing system to: collect data for processes operating across one or morea plurality of layers of an enterprise technology ecosystem, the data including logging and monitoring data fromone or more infrastructure ecosystemplatforms corresponding to the processes , wherein each layer of theplurality of layers comprises a distinct functional role within the enterprise technology ecosystem
  2. identify relationships among the collected data using artificial intelligence
  3. predict stability of theplurality of layers using a first machine learning model that analyzes the logging and monitoring data to assess stability across theplurality of lay ers : generate a stability score for the enterprise technology ecosystem based on the collected data using a second machine learning model that processes output from the first machine learning model, wherein the stability score indicates overall health of the enterprise technology ecosystem by a quantifiable measure
  4. predict imbalances within the enterprise technology ecosystem based on the relationships
  5. optimize the processes operating across the layers by: identifying one or more dependencies in the collected data, synthesizing the collected datatickets and epics using a generative Al model based on information from planning and program management tools based on the one or more dependenciesto determine whetherpredicted issues are already being addressed and, if not, to generate actionable tasks, and updating sources of the data in real-time based on event-triggered updates
  6. maintain operational equilibrium of the processes across all layers bygrecommendinimplementingproactive and reactive measures based on the collected data to address errors, issues, and remediation needs across the enterprise technologyecosystem. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20260010860A1 · filed 7/5/2024
• TIAA awaiting examination published 1/8/2026

  AI-based engine for generating information elements

  Describes an AI engine that takes a user prompt together with a user profile or a market, industry, or behavior report and generates console data and an information element in response, serving general information generation.

  How it works

  A user interaction module passes a prompt from the interface to a neural console, which feeds the prompt plus at least one source such as a user profile, market performance report or industry report into a generative model to produce console data, and the module turns that data into an information element displayed back to the user.

  Where it fits

  Customer-facing financial and retirement account guidance.

  Stage

  Filed July 2024, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computing system for generating information elements, comprising:

  1. one or more processors
  2. one or more memories having stored thereon: (i) an information element generative model
  3. (ii) a set of user profile computer-executable instructions that, when executed by the one or more processors, cause a user profile module to provide user profiles
  4. (iii) a set of market performance report computer-executable instructions that, when executed by the one or more processors, cause a market performance module to provide market performance reports
  5. (iv) a set of industry report computer-executable instructions that, when executed by the one or more processors, cause an industry report module to provide industry reports
  6. (v) a set of behavior report computer-executable instructions that, when executed by the one or more processors, cause a behavior report module to provide behavior reports
  7. (vi) a set of user interaction computer-executable instructions that, when executed by the one or more processors, cause a user interaction module to: generate a user interface
  8. receive a user prompt via the user interface
  9. transmit the user prompt to a neural console
  10. receive console data from the neural console
  11. generate an information element based on the console data
  12. present the information element via the user interface
  13. (v) a set of neural console computer-executable instructions, when executed by the one or more processors, cause the neural console to: receive at least one of (a) a user profile from the user profile module, (b) a market performance report from the market performance module, (c) an industry report from the industry report module, or (d) a behavior report from the behavior report module
  14. receive the user prompt from the user interaction module
  15. generate the console data via the information element generative model based on the user prompt and the at least one of the user profile, the market performance report, the industry report, or the behavior report
  16. transmit the console data to the user interaction module. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260010759A1 · filed 7/3/2024
• State Farm awaiting examination published 1/1/2026

  Augmented reality system to provide recommendation to repair or replace an existing device to improve home score

  Describes an augmented reality system that overlays recommended device updates onto a view of a structure and displays an improved home score, which would support property risk assessment and loss prevention.

  How it works

  The method takes camera data of a device near a structure, identifies the device, determines a recommended repair or replacement, calculates how that change would improve a home score, then overlays the recommendation and the score improvement onto the live view through an AR viewer.

  Where it fits

  Property underwriting and policyholder loss prevention.

  Stage

  Filed September 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computer-implemented method of using Augmented Reality (AR) for visualizing a recommended update to an existing device proximate a structure, the method comprising:

  1. receiving, with one or more processors, underlay layer data indicative of a field of view associated with an AR viewer device, wherein the field of view comprises the existing device proximate the structure
  2. identifying, by the one or more processors, the existing device in the underlay layer data
  3. determining, by the one or more processors, the recommended update to the existing device
  4. calculating, by the one or more processors, an improvement to a home score associated with the structure based upon the recommended update
  5. receiving, with the one or more processors, overlay layer data including an indication of the recommended update to the existing device and the improvement to the home score
  6. correlating, with the one or more processors, the overlay layer data with the underlay layer data
  7. creating, with the one or more processors, an AR display based upon the correlation, the AR display including the indication of the recommended update and the improvement to the home score
  8. displaying, with the one or more processors, the AR display to a user via the AR viewer device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260004361A1 · filed 9/3/2025
• State Farm under examination published 1/1/2026

  Systems and methods for generating user offerings responsive to telematics data

  Describes a device that builds a driver profile from vehicle telematics, combines it with contextual environment data, and generates offerings to influence driving behavior, supporting auto pricing and servicing.

  How it works

  From a driver profile built on historical telematics the device selects an offering and a driving-behavior recommendation carrying an incentive, displays them to the driver, then checks current trip data to see whether the recommendation was followed and sends a confirmation that the incentive was earned when it was.

  Where it fits

  Usage-based auto insurance pricing and policyholder engagement.

  Stage

  Filed September 2025, under examination, written from the claim as originally filed.

  The claim as filed

  An analytics computing device for processing vehicle-based telematics data, the analytics computing device comprising at least one processor in communication with a memory device, the at least one processor configured to:

  1. based upon a driver profile of a driver created from historical telematics data of the driver, select a user offering and an indication of driving behavior of the driver
  2. cause to be displayed, on a user computing device associated with the driver, the user offering, the indication of the driving behavior of the driver and a recommendation including at least one of (i) one or more alternate forms of transportation for the driver to use on trips or (ii) actions to improve the driving behavior, the user offering including an incentive if the recommendation is implemented
  3. after the user offering is displayed: determine whether the recommendation has been implemented by analyzing current telematics data associated with current trips taken by the driver
  4. in response to determining that the recommendation has been implemented, transmit, to the user computing device, a confirmation message indicating that the driver has earned the incentive. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/11/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  References cited against it: US 10,198,879.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260004328A1 · filed 9/3/2025
• TIAA allowed published 1/1/2026

  Digital wallet applications supporting decentralized web integration

  Describes a digital wallet that draws data from a user's linked electronic accounts and applies a machine learning model to generate recommendations, shown in a wallet app integrated with decentralized web technology.

  How it works

  The system gathers data from sources tied to a set of electronic accounts linked to a user's digital wallet, runs it through at least one machine learning model to generate an account-related recommendation, and displays that recommendation inside a wallet application integrated with decentralized web technology.

  Where it fits

  Consumer financial and retirement account servicing; the abstract names no specific insurance workflow.

  Stage

  Filed September 2025, applicant has responded to the examiner, written from the claim as originally filed.

  The claim as filed

  A system comprising:

  1. one or more processors
  2. a non-transitory computer-readable memory coupled to the one or more processors, the non-transitory computer-readable memory storing instructions thereon that, when executed by the one or more processors, cause the one or more processors to: obtain data from a set of data sources, the data being associated with a set of electronic accounts linked to a digital wallet of a user
  3. generate, using at least one machine learning (ML) model based on the data, a recommendation for the user related to the set of electronic accounts
  4. display the recommendation in a digital wallet application that (i) implements the digital wallet and (ii) is integrated with a decentralized web technology. Original 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 3/9/2026. The grounds raised so far:

  • section 102, anticipation: the examiner found a single earlier reference showing the same thing

  The examiner has pointed at allowable subject matter in claims 3, 5, 6, 7, 8, 9, 10, 11, 12, 13, 16, 17, 18, 19. That is the examiner naming the limitation that would earn the patent, not a statement that it will be granted.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260004279A1 · filed 9/3/2025
• Prudential under examination published 1/1/2026

  Methods and systems for adaptive data trend prediction and visualization

  Proposes training a projection model on historical process and performance metrics to predict a project trend and visualize it alongside current metric readings as new data arrives.

  How it works

  An application tracks process and performance metrics for a project, trains a projection model by grouping historical indicators into fixed-length trend windows and learning from the windows that precede them, then applies that model to a window of live metrics to generate and chart a predicted performance trend alongside the current readings.

  Where it fits

  General project performance analytics; the abstract states no specific insurance workflow.

  Stage

  Filed June 2025, under examination, written from the claim as originally filed.

  The claim as filed

  A computer implemented method for real-time data prediction and visualization, comprising:

  1. executing an information management application for tracking a plurality of metrics associated with a project, the plurality of metrics including a set of process metrics and a set of performance metrics
  2. extracting, from a historical database, historical data of the plurality of metrics including a temporal series of historical metric indicators of each metric, each historical metric indicator corresponding to a respective sampling window having a respective temporallength
  3. generating current data including a temporal series of current metric indicators of each of the plurality of metrics, each current metric indicator corresponding to a respective sampling window having a respective temporal length
  4. identifying a target projection length
  5. training a performance projection model using the historical data, further including: grouping the temporal series of historical metric indicators of a subset of metrics to a plurality of metric indicator sets, each metric indicator set corresponding to a respective trend window having the target projection length
  6. for each of the plurality of metric indicator sets: determining a respective performance trend corresponding to the respective trend window for one or more first performance metrics
  7. using the respective performance trend as a ground truth
  8. identifying a subset of historical metric indicators, which is sampled in a respective prediction window that precedes at least a subset of the respective trend window
  9. training the performance projection model using the subset of historical metric indicators and the respective performance trend
  10. at a first time, while collecting the current data, identifying a subset of current metric indicators that corresponds to a current prediction window and includes a recent current indicator sampled immediately before or at the first time
  11. applying the performance projection model to process the subset of current metric indicators, thereby generating a predicted performance trend of one or more first performance metrics corresponding to a current trend window identified by the target projection length
  12. visualizing the predicted performance trend of the one or more first performance metrics jointly with the subset of current metric indicators. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 7/1/2026, most recent 8/10/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Bilicki et al, Achin et al, Rajnayak et al, Joglekar et al.

  A further rejection under section 101 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20260004222A1 · filed 6/24/2025
• State Farm awaiting examination published 1/1/2026

  Methods and systems for preparing unstructured data for statistical analysis using electronic characters

  Describes turning time-ordered process events into emoji sequences, converting those into feature vectors, and feeding them to a machine learning technique to prepare unstructured data for analysis.

  How it works

  The method identifies the time-ordered events that occurred during a process, encodes each event as a categorical value to form a sequence in that time order, and generates a graphical representation of the sequence; the broader disclosure uses such sequences as feature vectors for machine learning.

  Where it fits

  Preparing unstructured operational or claims process data for statistical analysis; no specific line is named.

  Stage

  Filed September 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A method for visualizing a process, the method comprising:

  1. Identifying, by one or more processors, a time-ordered sequence of events that occurred during the process
  2. generating, by the one or processors, a categorical value sequence, each categorical value in the categorical value sequence representing an event of the events that occurred during the process, the categorical value sequence being ordered in accordance with the time-ordered sequence
  3. generating, by the one or more processors, a graphical representation of the categorical value sequence. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20260004076A1 · filed 9/16/2025
• Hartford responded published 12/25/2025

  System and method for conversational generative AI driven underwriting assistant

  Describes a generative AI assistant that extracts and summarizes information from sources such as knowledge graphs, websites and loss reports to answer risk relationship queries, aimed at underwriting support.

  How it works

  A back-end server answers a risk relationship query by having a deep-learning and NLP extractor summarize information from sources such as knowledge graphs, websites and historical loss reports, then passing it to neuro-symbolic language model agents that generate a response carrying alerts and suggestions from the risk data store.

  Where it fits

  Commercial underwriting decision support.

  Stage

  Filed October 2024, applicant has responded and the claim has already been amended during prosecution.

  The claim as amended, showing changes

  A risk relationship analysis system implemented via a back-end application computer server of an enterprise, comprising:

  1. (a) a risk relationship data store that contains electronic records associated with a plurality of risk relationships between the enterprise and parties, and, for each risk relationship, a risk relationship identifier, a party identifier, and at least one risk relationship parameter
  2. (b) the back-end application computer server, coupled to the risk relationship data store, including: a computer processor, and a computer memory coupled to the computer processor and storing instructions that, when executed by the computer processor, cause the back-end application computer server to: receive a risk relationship analysis request from a user device, extract and summarize information, by a data extractor using deep learning and natural language processing from multiple data sources, including knowledge graphs, websites, and historical loss reports, receive, by a plurality of specialized Neuro-Symbolic Large Lan guag e Model (NS-LLM) agents, the extracted and summarized information, generate, by the NS-LLM agents using both neural and symbolic Artificial Intelligence approaches, a response generator, an accurate and contextually relevant response based on the extracted data and information in the risk relationship data store, the response including guidance associated with alerts and suggestions, and transmit the relevant response to the user device. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 4/20/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250390825A1 · filed 10/22/2024
• State Farm awaiting examination published 12/25/2025

  Generative artificial intelligence for a network security scanner

  Proposes prompting an AI model to write network security testing code, then running that code to scan a network for devices and vulnerabilities and report them to a user.

  How it works

  On receiving a security vulnerability announcement the system prompts an ML chatbot to generate testing code for that vulnerability, receives the code, and executes it to scan the network for devices, identify devices affected by the announced vulnerability, and report them to a user.

  Where it fits

  Internal cybersecurity operations; the abstract states no insurance-specific workflow.

  Stage

  Filed August 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computer system for network security vulnerability inspection, the computer system comprising:

  1. one or more processors
  2. a memory storing executable instructions thereon that, when executed by the one or more processors, cause the one or more processors to: receive a security vulnerability announcement, transmit a prompt for a network security vulnerability testing code and the security vulnerability announcement to a machine learning (ML) chatbot to cause an ML model to generate the network security vulnerability testing code, receive the network security vulnerability testing code from the ML chatbot, and alert a user regarding the security vulnerability announcement and/or the network security vulnerability testing code
  3. wherein the network security vulnerability testing code comprises further instructions that, when executed by the one or more processors, cause the one or more processors to: scan a network to identify network computing devices, scan one or more of the network computing devices to identify security vulnerabilities related to the security vulnerability announcement and vulnerable network computing devices, and communicate an identification of the security vulnerabilities and/or the vulnerable network computing devices to the user. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250390418A1 · filed 8/25/2025
• State Farm final rejection published 12/25/2025

  Artificial intelligence route generation and cross-platform mobile application for a gig ecosystem

  Describes a cross-platform app that displays trip assignments and, once a user selects them, generates an optimal route using an AI model trained on historical trip records.

  How it works

  The app displays trip assignments from multiple provider systems, takes the worker's selection of a set of assignments each carrying a trip type and an origin and destination, and feeds those into an AI model trained on historical records of concurrently performed trips to generate and display one optimal route across them.

  Where it fits

  Gig-economy driver platforms; the abstract names no specific insurance workflow.

  Stage

  Filed May 2025, applicant has responded and the claim has already been amended during prosecution.

  The claim as amended, showing changes

  A computing device configured to use an artificial intelligence (AI) model to generate routes for concurrently-performed trips associated with different trip types, the computing device comprising at least one processor in communication with at least one memory device and with a user device corresponding to a user, the at least one processor configured to:

  1. cause , usingan application executing on the user device , the user deviceto display a plurality of trip assignments, each of the plurality of trip assignments associated with a trip assignment provider device of a plurality of trip assignment provider devices
  2. receive, from the user device, a selection of a set of trip assignments one or moreof the plurality of trip assignments, each of the set of one or moreselected trip assignments including trip information defining at least a respective trip type, the trip information including at least an origin and a destination
  3. generate an optimal route for the set of selected trip assignments based upon an output of the AImodel generated in response to an input of the selected trip assignments including the trip information defining the respective trip type to the AI)Iusing an artificial intelligence (Amodel, wherein the Almodel is trained using historical trip records including historical trip information associated with historical trips, the historical trip information including historical trip types of historical trips that were performed concurrently
  4. and cause , usingthe application [[,] ] executing on the user device to display the generated optimal route for the set of selected trip assignments. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 2/26/2026, most recent 9/4/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250389542A1 · filed 5/19/2025
• State Farm awaiting examination published 12/18/2025

  Systems and methods for artificial intelligence based reinforcement training and workflow management for one or more chatbots

  Describes an AI module that transcribes a spoken statement, infers its intent, recommends and analyzes a response, and produces a spoken reply delivered through a selected chatbot, aimed at contact handling.

  How it works

  An orchestration device receives a customer statement, determines its intent, selects one or more chatbots from a pool based on that intent to start an audio conversation, while an AI module monitors the exchange and drives a representative-facing interface that displays the conversation data.

  Where it fits

  Customer service and claims call handling.

  Stage

  Filed May 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computer system for applying chatbots and artificial intelligence (AI ) tools to automatically respond to a submitted statement and generate a representative interface to monitor the response, the computer system comprising:

  1. a plurality of chatbots
  2. an orchestration computing device comprising at least one first processor in communication with at least one first memory device, and further in communication with the plurality of chatbots and a user computer device associated with a user
  3. an AImodule comprising at least one second processor in communication with at least one second memory device, and further in communication with the orchestration computing device, wherein the at least one first processor of the orchestration computing device is programmed to: receive, from the user computing device, a statement of the user
  4. determine at least one intent of the statement
  5. select one or more chatbots from the plurality of chatbots to analyze the statement based upon the at least one intent
  6. initiate an audio conversation with the user using the selected one or more chatbots, and wherein the at least one second processor of the AImodule is programmed to: monitor the audio conversation between the user and the selected one or more chatbots
  7. cause the representative interface to be displayed on a representative computing device associated with a representative that includes data representing the audio conversation between the user and the one or more chatbots. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250385880A1 · filed 5/23/2025
• State Farm awaiting examination published 12/18/2025

  Systems and methods for artificial intelligence based reinforcement training and workflow management for one or more chatbots

  Describes an AI module that transcribes a spoken statement, infers its intent, recommends and analyzes a response, and produces a spoken reply delivered through a selected chatbot, aimed at contact handling.

  How it works

  The system receives a caller's spoken statement, detects pauses to divide it into separate utterances, identifies an intent for each, selects a chatbot per utterance based on its intent, and generates a single audio response from the combined outputs of the selected chatbots.

  Where it fits

  Automated customer service and call handling.

  Stage

  Filed May 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computer system for controlling a plurality of chatbots and artificial intelligence (AI) tools used to respond to a submitted statement by a caller, the computer system comprising:

  1. an orchestration computing device comprising at least one first processor in communication with at least first one memory device, and further in communication with the plurality of chatbots and a user computing device associated with the caller
  2. an AI module comprising at least one second processor in communication with at least one second memory device, and further in communication with the orchestration computing device, the at least one first processor of the orchestration computing device programmed to: receive, from the user computing device, a verbal statement of the caller including a plurality of words
  3. detect one or more pauses in the verbal statement
  4. divide the verbal statement into a plurality of utterances based upon the one or more pauses and input from the AI module
  5. identify, for each of the plurality of utterances, an intent
  6. select, for each of the plurality of utterances, based upon the intent of the corresponding utterance, a chatbot to analyze the utterance of the plurality of utterances
  7. generate an audio response from an output from each of the selected chatbots, the audio response responsive to the verbal statement. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250384877A1 · filed 5/23/2025
• State Farm awaiting examination published 12/18/2025

  Systems and methods for artificial intelligence based reinforcement training and workflow management for one or more chatbots

  Describes an AI module that transcribes a spoken statement, infers its intent, recommends and analyzes a response, and produces a spoken reply delivered through a selected chatbot, aimed at contact handling.

  How it works

  An AI module transcribes a user's spoken statement to text, augments it by determining its intent, produces recommendations for a response, analyzes the augmented text against those recommendations, then generates audio-response data and plays it back by having a selected chatbot execute it.

  Where it fits

  Automated customer service and call handling.

  Stage

  Filed May 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computer system for training a plurality of chatbots using artificial intelligence (AI) tools to process statements, the computer system comprising:

  1. an orchestration computing device comprising at least one first processor in communication with at least one first memory device, and further in communication with a plurality of chatbots and a user computer device associated with a user
  2. an Almodule comprising at least one second processor in communication with at least one second memory device, and further in communication with the orchestration computing device, the at least one second processor of the Almodule programmed to: receive, from the user computer device via the orchestration computing device, a verbal statement of the user including a plurality of words
  3. translate the verbal statement into a text statement
  4. augment the text statement by determining at least one intent of the text statement
  5. provide recommendations for responding to the augmented text statement
  6. analyze the augmented text statement and the recommendations
  7. generate data representing an audio response to the analyzed augmented text statement
  8. present the audio response to the user by causing a selected chatbot to execute the generated data. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250384875A1 · filed 5/23/2025
• State Farm under examination published 12/18/2025

  Artificial intelligence-powered building inspection system

  Proposes generating a guided building assessment plan, collecting a user's unstructured observations, and using an AI model to compose a structured inspection report from building data and those observations.

  How it works

  The system takes building data with structural information, determines a layout, generates a guided plan posing questions about the structure to collect a user's unstructured observations, then feeds the data, observations and responses into an AI model that generates a structured inspection report.

  Where it fits

  Residential property inspection and underwriting.

  Stage

  Filed July 2024, under examination, written from a claim the applicant has already amended during prosecution.

  The claim as amended, showing changes

  A building inspection system for generating a structured building inspection report for at least a portion of a residential building, the building inspection system comprising:

  1. one or more memory devices having instructions stored thereon that, when executed by one or more processors, cause the one or more processors to perform operations comprising: receiving building data for the residential building from one or more data sources, wherein the building data includes structural information about the at least a portion of the residential building
  2. determining a layout for the at least a portion of the residential building based upon the structural information
  3. generating a guided building assessment plan on a user device based upon the building data, wherein the guided building assessment plan provides inspection instructions for a user to gather unstructured user observations of the at least a portion of the residential building, and wherein the guided building assessment plan comprises one or morequestions provided to the user regardingone or more pieces of the structural information
  4. receiving the unstructured user observations
  5. automatically generating, using an artificial intelligence model, the structured building inspection report in a predetermined format for delivery to one or more users associated with the residential building based upon the building data [[and]] the unstructured user observations, and one or more responses from the user to the one or morequestions. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 9/11/2025, most recent 6/2/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Graham et al, Brown et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250384497A1 · filed 7/22/2024
• Allstate under examination published 12/18/2025

  Consumer engagement and management platform using machine learning for intent driven orchestration

  Describes training models on historical data to identify a customer's intent, generate an engagement output, and select a communication channel to deliver it, aimed at customer engagement.

  How it works

  The platform reads an individual's intent with a selected model, picks output-generation models to produce a customer engagement message, uses channel models to choose the channel most likely to provoke a response, has an enterprise device format and display it, and retrains itself on new interaction data seeded from a labelled set.

  Where it fits

  Customer engagement, marketing and retention.

  Stage

  Filed May 2025, under examination, written from a new claim substituted after the original claims were cancelled.

  The claim new claim

  A computing platform, comprising:

  1. at least one processor
  2. a communication interface communicatively coupled to the at least one processor
  3. memory storing computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: identify, using a selected one of the plurality of intent identification models, an intent of an individual
  4. select, based on the intent of the individual, one or more engagement output generation models
  5. generate, using the selected one or more engagement output generation models, a customer engagement output
  6. identify, using one or more communication channel models, a communication channel, wherein the one or more communication channel models identify the communication channel by analyzing the intent of the individual to determine the communication channel that will provoke the individual to engage with the customer engagement output
  7. send one or more commands directing an enterprise user device to format the customer engagement output based on the communication channel to generate a communication channel format for the customer engagement output and display the customer engagement output on a graphical user interface associated with the communication channel, the customer engagement output generated using the selected one or more engagement output generation models selected based on the intent of the individual identified using the selected one of the plurality of intent identification models, wherein sending the one or more commands directing the enterprise user device to display the customer engagement output causes the enterprise user device to display the customer engagement output using the communication channel and on the graphical user interface associated with the communication channel in the communication channel format
  8. continuously train the one or more intent orchestration models based on post-historical data comprising the identified intent and real-time data corresponding to the individual, wherein training of the one or more intent orchestration models comprises training one or more supervised learning models to automatically assemble a labelled dataset of historical data by initially inputting a manually labelled dataset into one or more intent orchestration models comprising the plurality of intent identification models and automatically generating the labelled datas

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 6/25/2026. The grounds raised so far:

  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250384397A1 · filed 5/23/2025
• Equifax awaiting examination published 12/18/2025

  Clustering techniques for machine learning models

  Describes a two-pass clustering method for large datasets that would group records, refine the groups using derived cluster features, and select training samples from the result to train a neural network that computes a risk indicator for an entity.

  How it works

  The method would first split a dataset into a chosen number of clusters, derive special features for those clusters, then re-cluster on those features; a neural network is trained on samples drawn from the refined clusters and applied to an entity's predictor variables to compute a risk indicator in response to a query.

  Where it fits

  Building and querying risk scoring models over large populations; general risk assessment rather than a named insurance line.

  Stage

  Filed June 2025 and published December 2025; pending examination on the originally filed claim, not yet narrowed during prosecution.

  The claim as filed

  A method that includes one or more processing devices performing operations comprising:

  1. clustering a dataset into a set of clusters, wherein the clustering comprises: determining a number of clusters to be generated for the dataset
  2. clustering the dataset into the determined number of clusters
  3. determining a plurality of special features for the determined number of clusters
  4. re-clustering the dataset based on the plurality of special features to generate the set of clusters
  5. training a neural network model for computing a risk indicator from predictor variables based on the set of clusters wherein the neural network model is trained based on training samples selected from the set of clusters, the training samples comprising training predictor variables and training outputs corresponding to the training predictor variables
  6. receiving, from a remote computing device, a risk assessment query for a target entity
  7. computing, responsive to the risk assessment query, an output risk indicator for the target entity by applying the trained neural network model to predictor variables associated with the target entity
  8. transmitting, to the remote computing device, a responsive message including the output risk indicator, wherein the output risk indicator is usable for controlling access to one or more interactive computing environments by the target entity. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250384267A1 · filed 6/20/2025
• Allstate awaiting examination published 12/18/2025

  System and method for simulating traffic using agent-based modeling

  Proposes feeding real-world agent position data into a machine learning simulation that outputs synthetic traffic movement metrics for a chosen geographic area and time frame.

  How it works

  The system feeds real-world agent-based position data for a selected geographic area and timeframe into a machine learning model, generates a simulated environment with movement constraints and a variation of it, and outputs metrics describing synthetic driver movement over time on a map, including metrics for the varied scenario.

  Where it fits

  Auto risk modeling and scenario analysis.

  Stage

  Filed June 2024, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A system comprising:

  1. one or more processors
  2. one or more machine-learning models of a simulation system, wherein the one or more machine-learning models are associated with custom selections of geographic areas and timeframes
  3. one or more memory units storing computer-executable instructions, which when executed by the one or more processors, cause the system to: input, in one of the machine-learning models, real-world agent-based position data associated with one of the custom selections of one of the geographic areas and one of the timeframes
  4. generate a simulated environment including one or more movement constraints that represent the one of the geographic areas and a variation of the simulated environment based on one or more changes to the simulated environment
  5. output, from the one of the machine-learning models, one or more metrics associated with synthetic agent-based position data over the one of the timeframes within a map for the simulated environment and one or more variation metrics associated with the variation of the simulated environment, wherein the one or more metrics represent synthetic movement behavior of one or more synthetic agents associated with one or more synthetic individuals based on real movement behavior associated with the geographic area and the timeframe, and wherein the one or more variation metrics represent synthetic movement behavior of the one or more synthetic agents associated with the one or more synthetic individuals in the variation of the simulated environment. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250384186A1 · filed 6/17/2024
• Allstate allowed published 12/11/2025

  Roadside assistance detection

  Describes collecting device data such as location, road type and ambient noise to detect a stopped vehicle and estimate the likelihood it needs roadside assistance.

  How it works

  The method detects when a vehicle moves from a non-highway to a highway road and raises sensor sampling, then compares its speed against a threshold and against the aggregate speed of nearby vehicles to calculate an emergency probability, and when that probability crosses a threshold it initiates roadside assistance or contacts the driver.

  Where it fits

  Roadside assistance dispatch and auto claims servicing.

  Stage

  Filed May 2025 and allowed but not yet granted, written from a new claim substituted after the original claims were cancelled.

  The claim new claim

  A method performed by a first computing device, the method comprising:

  1. obtaining a first set of data from one or more sensors associated with a vehicle
  2. determining, based on the first set of data, that a type of road travelled by the vehicle has changed from a non-highway road to a highway road
  3. in response to the determination that the type of road travelled by the vehicle has changed from the non-highway road to the highway road, causing a data collection rate to be increased for the one or more sensors associated with the vehicle, wherein the data collection rate is caused to be increased, at least temporarily, for the one or more sensors associated with the vehicle each time the computing device determines that the type of road travelled by the vehicle has changed from any non-highway road to any highway
  4. obtaining a second set of data from the one or more sensors associated with the vehicle
  5. determining, based on the second set of data, a vehicle speed of the vehicle
  6. determining, based on a comparison of the vehicle speed to a threshold speed, that the vehicle speed is equal to or below the threshold speed for a time period
  7. determining an aggregate vehicle speed for a plurality of other vehicles within a proximity to the vehicle
  8. calculating, based on the aggregate vehicle speed and further based on the determination that the vehicle speed is equal to or below the threshold speed for the time period, a probability that an emergency associated with the vehicle has occurred
  9. determining that the calculated probability exceeds a threshold probability
  10. following the determination that the calculated probability exceeds the threshold probability, either : (i) initiating roadside assistance for the vehicle
  11. or (ii) establishing communication with a driver and/or passenger of the vehicle that facilitates initiation of roadside assistance for the vehicle. New 3

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250378717A1 · filed 5/12/2025
• Assurant awaiting examination published 12/11/2025

  Apparatuses, computer-implemented methods, and computer program products for improved selection and provision of operational support data objects

  Proposes using data models to select the highest-confidence support data object for resolving a malfunction on a networked home device, seeking to raise first-time resolution rates.

  How it works

  While a transmission request runs behind a waiting-message panel, the apparatus renders in a second panel a predicted operational support data object, selected as the highest-confidence resource for the malfunction, as a link that retrieves a third-party support resource, and selecting that link terminates the underlying transmission process.

  Where it fits

  Connected-device and warranty support servicing.

  Stage

  Filed April 2025, awaiting examination, written from a new claim substituted after the original claims were cancelled.

  The claim new claim

  1. An apparatus comprising at least one processor and at least one memory having computer-coded instructions stored thereon, wherein the computer-coded instructions in execution with the at least one processor cause the apparatus to: cause rendering, to a requesting client device, of a user interface comprising a first sub- interface and a second sub-interface, wherein the first sub-interface comprises a waiting message interface element corresponding to a transmission process
  2. output a predicted operational support data object to the requesting client device by rendering, in the second sub-interface, an operational support interface element corresponding to the predicted operational support data object, wherein the predicted operational support data object comprises a link that, upon selection, causes the requesting client device to retrieve a third-party operational support data object corresponding to the link
  3. in response to the selection of the link associated with the predicted operational support data object, terminate the transmission process. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250377917A1 · filed 4/22/2025
• State Farm awaiting examination published 12/4/2025

  Systems and methods for parsing multiple intents in natural language speech

  Describes transcribing a spoken statement, labeling its words, detecting splits, dividing it into multiple intents, and generating a response, aimed at contact handling.

  How it works

  The system receives a spoken statement, builds a constituency tree of its grammar, detects split points by parsing that tree against grammar rules, divides the statement into separate utterances at those splits, identifies an intent for each utterance, and generates a response based on the set of intents.

  Where it fits

  Automated customer service and call routing.

  Stage

  Filed August 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  1. A computer system for generating responses to a verbal input by parsing separate intents in natural language speech, the computer system comprising at least one processor in communication with at least one memory device, the at least one processor is programmed to: receive a verbal statement of a user including a plurality of words
  2. generate a constituency tree structure based upon the verbal statement
  3. detect one or more splits in the verbal statement by parsing the constituency tree structure based upon at least one of a plurality of grammar-related rules
  4. divide the verbal statement into a plurality of utterances based upon the one or more splits
  5. analyze each of the plurality of utterances to identify a plurality of intents, wherein each of the plurality of intents corresponds to one of the plurality of utterances
  6. generate a response based on the plurality of intents. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250372087A1 · filed 8/13/2025
• State Farm allowed published 12/4/2025

  Artificial Intelligence for Flood Monitoring and Insurance Claim Filing

  Proposes detecting a flood event in a structure and directing an ML chatbot to file a flood reimbursement claim with the insurer by phone, converting text output into voice.

  How it works

  The system detects a flood event in a structure and transmits information about the structure and a prompt to a machine learning chatbot, causing the chatbot to request flood remediation services over a telephone call by converting its text output into a voice output.

  Where it fits

  Flood claims handling and loss mitigation.

  Stage

  Filed August 2025 and allowed but not yet granted, written from the claim as originally filed.

  The claim as filed

  A computer system for improved flood monitoring and flood remediation using a machine learning chatbot to facilitate interaction with a flood remediation service provider, the computer system comprising:

  1. one or more processors
  2. a memory storing executable instructions thereon that, when executed by the one or more processors, cause the one or more processors to: detect a flood event in a structure, and transmit information associated with the structure and a prompt for flood remediation services to a machine learning (ML) chatbot to cause the ML chatbot to request flood remediation services for the flood event via telephone by converting a text output into a voice output. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250371632A1 · filed 8/22/2025
• Assurant awaiting examination published 12/4/2025

  Systems, methods, and apparatuses for predictive performance analysis

  Describes applying unit performance data to trained models to produce a predictive performance data set for an analytical unit and rendering it as virtual widgets on a device screen.

  How it works

  The method receives performance data for an analytical unit, applies it to one or more trained performance-analysis models to generate a predictive performance data set, builds renderable virtual widgets representing that data, and displays them on a user device screen.

  Where it fits

  General performance analytics; the abstract states no specific insurance workflow.

  Stage

  Filed May 2025, awaiting examination, written from the claim as originally filed.

  The claim as filed

  A computer-implemented method comprising:

  1. receiving, by one or more processors and from one or more data sources, unit performance data for an analytical unit
  2. applying, by the one or more processors, the unit performance data to one or more trained performance analysis models to generate a predictive performance data set for the analytical unit by analyzing the unit performance data using the one or more trained performance analysis models
  3. generating, by the one or more processors, one or more renderable virtual widgets comprising one or more representations of at least a portion of the predictive performance data set for the analytical unit
  4. displaying, by the one or more processors, the one or more renderable virtual widgets on a screen of a user device. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250370776A1 · filed 5/30/2025
• EXL allowed published 12/4/2025

  Machine learning based code migration engine

  Proposes chunking legacy code into prompts for a neural network to generate replacement code, adjusting input chunk and output window sizes to limit truncation during migration.

  How it works

  The engine breaks source-language code into normalized chunks kept under a neural network's token limit, builds a prompt that preserves code and data dependencies, applies a trained converter to produce target-language code, and when the output exceeds a size threshold it lowers it and re-prompts on a reselected subset to avoid truncation.

  Where it fits

  Software modernization and developer tooling; the abstract states no insurance-specific use.

  Stage

  Filed May 2024 and allowed but not yet granted, written from the claim as originally filed.

  The claim as filed

  1. At least one non-transitory, computer-readable storage medium comprising instructions recorded thereon, the instructions, when executed by at least one processor of a code generator, causing the code generator to perform migration of computer code from a computer language to a different computer language by: using a first code unit in a source language, generating a set of input chunks, wherein a size of the input chunks is normalized across the set of input chunks, and wherein the size of the input chunks does not exceed a predetermined token limit for a trained code converter neural network
  2. using a first subset of input chunks from the set of input chunks, generating a first prompt for the trained code converter neural network, wherein the first prompt is structured to preserve at least one of a code dependency, data dependency, code type, or sequence dependency in the first code unit
  3. applying the trained code converter neural network to the first subset of input chunks from the set of input chunks to generate an output item that comprises a second code unit in a target language, wherein the trained code converter neural network is configured to generate output in conformance to a size threshold
  4. in response to a determination that a size of the output item exceeds the size threshold, dynamically reducing the size threshold and generating a second prompt for the trained code converter neural network, wherein the second prompt is structured to operate on a second subset of input chunks of size M selected at least in part from a first subset of input chunks of size N, wherein N < M
  5. applying the trained code converter neural network to the second subset of input chunks to generate a third code unit
  6. causing a computing device to display the generated second code unit or third code unit. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250370725A1 · filed 5/31/2024
• Elevance under examination published 11/27/2025

  Systems and Methods for Authorization Automation Using Artificial Intelligence

  Describes using optical character recognition to extract text from a historical record and applying a natural language model against authorization guidelines to decide whether a treatment meets approval criteria.

  How it works

  The method OCRs the patient's historical record into text, decides whether the treatment has guidelines to analyze by machine learning, then identifies the authorization criteria and a matching NLP model, runs that model over the extracted text to find relevant record data, and authorizes the treatment when the data meets the criteria.

  Where it fits

  Health insurance prior authorization and utilization management.

  Stage

  Filed April 2025, under examination, written from the claim as originally filed.

  The claim as filed

  1. A machine learning based method for authorizing the performance of a treatment, comprising the steps ofreceiving a treatment authorization request for a treatment, the treatment authorization request including a historical record of the person who will receive the treatment and treatment identifying information relating to the treatment
  2. creating an extracted text of the historical record using optical character recognition on the historical record
  3. determining whether to analyze authorization performance of the treatment using a machine learning authorization process, wherein the determination is based on treatment identifying information and whether treatment authorization guidelines exist for the treatment
  4. in response to a determination to analyze authorization performance of the treatment using a machine learning authorization process: identifying authorization criteria for the treatment based on the treatment authorization guidelines, wherein the authorization criteria includes records data conditional to authorization of performance of the treatment
  5. identifying a natural language record processing model corresponding to the treatment authorization guidelines
  6. performing natural language processing on the extracted text of the record in accordance with the identified natural language record processing model to identify relevant record data in the record
  7. determining whether the relevant record data meets the authorization criteria
  8. in response to a determination that the relevant record data meets the authorization criteria, authorizing the treatment. 058440-03-5081-US O1

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 6/3/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250364138A1 · filed 4/18/2025
• State Farm responded published 11/27/2025

  Systems and methods for modeling telematics, positioning, and environmental data

  Proposes building a model relating historical liability amounts to personal, telematics and environmental data, then applying it to a candidate user's current data, relevant to auto risk assessment.

  How it works

  The device builds a machine learning model from training datasets of historical vehicle driving data, collects current data for a candidate driver, folds it into updated training datasets to continuously retrain the model, and runs the updated model to output a current coverage level for the driver including a likelihood of an accident.

  Where it fits

  Auto insurance underwriting and pricing.

  Stage

  Filed August 2025, applicant has responded and the claim has already been amended during prosecution.

  The claim as amended, showing changes

  A computing device for building artificial intelligence models, the computing device comprising at least one processor in communication with at least one memory device, the at least one processor configured to:

  1. create a first plurality of training datasets including historical vehicle driving- related data
  2. for buildingbuild a computer model using the first plurality of training datasets and one or more machine learning programs
  3. collect current user data of a candidate user associated with driving a vehicle
  4. create a second plurality of training datasets by updating the first plurality of training datasets to include the collected current user data
  5. continuously update the computer model by applyinginputting, into the computer model, at least one of the second plurality of training datasets to the modelor additional plurality of training data sets created by the computing device
  6. and execute the updated computer model to determineoutput, from the updated computer model, a determination of a current coverage level for the candidate user including a likelihood of an accident involving the vehicle. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 4/29/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250363566A1 · filed 8/7/2025
• State Farm awaiting examination published 11/20/2025

  Thick client and common queuing framework for contact center environment

  Describes a browser-based thick-client framework for a contact center that distributes processes across a device and provides shared queuing and communication services with an integrated model implementation.

  How it works

  The client app signals availability, the system checks the status of two candidate service providers, selects one based on those statuses, sends it a request, and processes the returned data object back through the client.

  Where it fits

  Contact center servicing and communication routing, applicable to insurer call-center operations.

  Stage

  Filed July 2025 and in examination; the claim read is the original as filed.

  The claim as filed

  A method, comprising:

  1. receiving an indication of availability from a client application executing on a client device associated with an online communication platform
  2. determining a first status associated with a first service provider associated with the online communication platform
  3. determining a second status associated with a second service provider associated with the online communication platform
  4. determining a selected service provider, based at least in part on the first status and the second status, wherein the selected service provider comprises the first service provider or the second service provider
  5. transmitting a request to the selected service provider
  6. receiving a data object from the selected service provider based at least in part on the request
  7. processing the data object via the client application. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250358339A1 · filed 7/28/2025
• State Farm under examination published 11/20/2025

  Methods and apparatus for automated claim processing using historical data

  Proposes estimating vehicle damage from images by finding similar damaged vehicles, scoring the likelihood a part is damaged, and deciding whether to include it in a repair estimate.

  How it works

  From a damage image and the vehicle type, the method picks a matching machine learning model, finds stored images of comparable damaged vehicles, estimates the probability a given component is damaged, and when that probability clears a threshold computes a repair or replacement cost.

  Where it fits

  Automated auto physical damage claims estimation and adjuster support.

  Stage

  Filed July 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A method of estimating damage to a vehicle, the method comprising:

  1. receiving, by a processor, an image illustrating damage to a vehicle, the vehicle being characterized by a vehicle type
  2. selecting, by the processor and based on the vehicle type, a machine learning algorithm configured to identify similarities between digital images illustrating damaged vehicles of the vehicle type
  3. identifying, by the processor, and using the machine learning algorithm and the image, a plurality of stored images of damaged vehicles of the vehicle type, wherein the plurality of stored images show vehicle damage corresponding to the damage illustrated in the image
  4. identifying, by the processor and based on stored information associated with the plurality of stored images, a likelihood that the vehicle has a particular damaged vehicle component
  5. determining, by the processor, that the likelihood is greater than a threshold value
  6. determining, by the processor and based on the likelihood being greater than the threshold value, an estimated cost associated with repair or replacement of the particular damaged vehicle component. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/6/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  References cited against it: Hanson et al, Li et al, Hanson.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250356428A1 · filed 7/24/2025
• Unum awaiting examination published 11/20/2025

  AI hallucination and jailbreaking prevention framework

  Describes chaining three large language models, one to screen and transform the prompt, one to generate, and one to check the output for hallucinations and harmful content.

  How it works

  A first language model receives the user prompt and produces a revised prompt, a second model generates a response from that revised prompt, and a third model reviews and where needed rewrites that response to remove inaccuracy or harmful content before it reaches the user.

  Where it fits

  General AI safety and output control; the abstract states no specific insurance workflow.

  Stage

  Filed August 2025; the original claims were cancelled and this claim substituted, typically following a rejection.

  The claim new claim

  A computer system configured to provide a generative artificial intelligence (A l) framework having multiple interconnected large language models, the computer system comprising:

  1. one or more physical processors
  2. one or more network interfaces configured to receive a user prompt from a user
  3. a memory configured to store one or more computer-readable instructions that, when executed by the one or more physical processors, configure the computer system to implement the generative Alframework, the generative Alframework comprising: a first processing stage comprising a first large language model, wherein the first large language model is configured to process the user prompt received at the one or more network interfaces and generate an updated user prompt, wherein the first large language model is configured to use a first machine learning model to generate the updated user prompt
  4. a second processing stage comprising a second large language model, wherein the second large language model is configured to process the updated user prompt generated by the first large language model and generate a response to the updated user prompt, wherein the second large language model is configured to use a second machine learning model to generate the response to the updated user prompt
  5. a third processing stage comprising a third large language model, wherein the third large language model is configured to process the response to the updated user prompt generated by the second large language model and generate a response to return to the user, wherein the third large language model is configured to use a third machine learning model to generate the response to return to the user, wherein at least one of the first processing stage, second processing stage, or third processing stage comprises an additional large language model. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250356139A1 · filed 8/6/2025
• Allstate under examination published 11/20/2025

  Intelligent vehicle notification systems and processes

  Proposes generating driver notifications through audio, visual or tactile channels from vehicle sensor and external data about road conditions, hazards, points of interest and related insurance offers.

  How it works

  The system ingests evacuation data and geographic evacuation-zone data from external sources, uses onboard sensors to determine that the vehicle qualifies as eligible within that zone, and transmits a notification of available evacuation funds to the driver.

  Where it fits

  Policyholder servicing and catastrophe-response communication for auto or property insurers.

  Stage

  Filed May 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A vehicle notification system, the vehicle notification system comprising:

  1. a notification feature communicatively coupled to a vehicle to notify a driver of the vehicle of a notification via the vehicle notification system
  2. one or more vehicle sensors associated with the vehicle
  3. one or more external data sources remote from the vehicle
  4. at least one processor communicatively coupled to the notification feature, the one or more vehicle sensors, and the one or more external data sources
  5. a memory communicatively coupled to the at least one processor
  6. one or more machine readable instructions stored in the memory that cause the vehicle notification system to perform at least the following when executed by the at least one processor: receive evacuation data and geographical evacuation zone data for a geographical evacuation zone from the one or more external data sources
  7. determine as an eligibility determination that the vehicle is one of one or more eligible vehicles within the geographical evacuation zone based on the one or more vehicle sensors
  8. transmit the notification to the driver of the vehicle via the notification feature of available evacuation funds based on the eligibility determination. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/24/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Foladare et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250353519A1 · filed 5/13/2025
• State Farm under examination published 11/13/2025

  Systems and methods for analyzing and mitigating community-associated risks

  Describes using a trained model to determine potential risks for a building from sensor and database data, then generating a risk profile and mitigation alerts or recommendations.

  How it works

  In the claim, the system builds a 3D electronic model of part of a city that includes planned future construction, runs a virtual simulation of a weather event across it, determines a potential risk to a building from that simulation, and updates the model to mark the risk at that building's location.

  Where it fits

  Property underwriting, catastrophe risk assessment and loss mitigation.

  Stage

  Filed July 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  1. A computer system for predicting impacts of new buildings by executing virtual simulations on virtual 3D models, the computer system comprising at least one processor in communication with at least one memory, the at least one processor programmed to: generate a 3D electronic model of at least part of a city, wherein the 3D electronic model comprises virtual electronic representations of a plurality of buildings including planned future construction of one or more buildings
  2. execute a virtual simulation on the 3D electronic model, the virtual simulation comprising at least one virtual weather event occurring within the city
  3. determine, based upon the virtual simulation, at least one potential risk associated with at least one building of the plurality of buildings
  4. cause the 3D electronic model to be updated to include a representation of the at least one potential risk at a location associated with the at least one building. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 6/16/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Pearson et al, Pearson, Pourmohammad et al, Tohidi et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250348967A1 · filed 7/22/2025
• Zesty.ai awaiting examination published 11/13/2025

  Hail Frequency Predictions Using Artificial Intelligence

  Describes training and validating a hail damage frequency model from properties that did and did not experience hail damage, using each property's location and features, to support property pricing and catastrophe risk assessment.

  How it works

  The method identifies a set of properties, selects a reduced subset of their available features, trains a hail damage frequency model on that subset, validates it, compares it against other model instances trained on different feature subsets, and selects the best-performing subset and model for production.

  Where it fits

  Property catastrophe pricing and underwriting for hail-exposed risks.

  Stage

  Filed July 2025 and in examination; the claim read has already been narrowed by amendment during prosecution, not granted.

  The claim as amended, showing changes

  A computer implemented method comprising:

  1. identifying, using one or more processors, a plurality of properties including a first set of properties that experienced hail damage and a second set of properties that did not experience hail damage
  2. determining, using the one or more processors, a location anda first subset of property features, wherein the first subset of property features is a reduced set selected from a set of available property features associated with for each property inthe plurality of properties
  3. training, using the one or more processors, a first damage frequency model based on the first subset of property features
  4. and validating, using the one or more processors, the first damage frequency model based on the first subset of property features comparing, using the one or more processors, the first damage frequencymodel based on the first subset of property features to other trained instances of the damage frequency model trained usingdifferent subsets of property features
  5. selecting, using the one or more processors, the first subset of features and the first damage frequency model for use in production based on the comparison. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250347824A1 · filed 7/11/2025
• Experian under examination published 11/6/2025

  Automatic data segmentation system

  Proposes predicting a debtor's recovery value from receivables, payment history and credit data, then assigning the individual to a collection segment to guide recovery strategy and benchmark collection efforts across clients.

  How it works

  Input data on an individual is fed to a hyperdimensional model trained on historical inputs and actual recovery values; the model returns a predicted recovery value, the individual is assigned to a segment sent to the client, and the later actual recovery is fed back to retrain the model.

  Where it fits

  Debt collection prioritization; adjacent to insurer subrogation and recovery operations.

  Stage

  Filed May 2025; the original claims were cancelled and this claim substituted in their place, typically after a rejection.

  The claim new claim

  A system for automatic data segmentation, the system comprising:

  1. one or more processors
  2. a non-transitory computer readable medium having stored thereon instructions that, when executed by the one or more processors, cause the one or more processors to: receive, from a client system, input data associated with a first individual of a first plurality of individuals for whom a client system has provided a service
  3. input, into a hyperdimensional model, the input data, wherein the hyperdimensional model has been trained by a training engine to generate predicted recovery values by: collecting historical data from the client system, wherein the historical data comprises historical input data and a plurality of actual recovery values, the historical input data and the plurality of actual recovery values corresponding to a second plurality of individuals for whom the client system has previously provided a service, generating first training data based on the historical data, wherein the hyperdimensional model comprises a plurality of dimensions, and wherein each dimension in the plurality of dimensions corresponds to a variable of the first training data, and training the hyperdimensional model using the first training data
  4. receive, from the hyperdimensional model, a predicted recovery value for the first individual
  5. based on the predicted recovery value, assign the first individual to a first segment of a plurality of segments
  6. transmit, to the client system, an electronic message comprising the first segment
  7. receive, from the client system a first actual recovery value for the first individual
  8. update the plurality of actual recovery values to include the first actual recovery value
  9. provide the updated plurality of actual recovery values to the training engine for additional training, retraining, or updating of the hyperdimensional model. New 22

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/15/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250342526A1 · filed 5/22/2025
• State Farm awaiting examination published 11/6/2025

  Simplistic machine learning model generation tool for predictive data analytics

  Describes a guided interface that walks a user through selecting data, choosing an algorithm and training a machine learning model, then distributing the trained model from a cloud server to client devices for predictive analytics.

  How it works

  Through a guided interface the tool takes a user's inputs to pull a dataset from a database, determine a correlated subset tied to a chosen data object, select a machine learning algorithm, build a model generation tool for that object, and deploy the tool on a server.

  Where it fits

  General predictive-analytics model building; no insurance-specific workflow is stated.

  Stage

  Filed July 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A method implemented by a computing device, the method comprising:

  1. generating a guided user interface (GUI) on the computing device
  2. receiving inputs, via the GUI, indicative of requested operations to be perf ormed by the computing device, the requested operations including: obtaining, from a database, a dataset associated with a plurality of data objects
  3. determining, based on a correlation associated with the plurality of data objects, a subset of the dataset associated with a first data object
  4. selecting a machine learning algorithm
  5. generating a machine learning (ML) model generation tool associated with the first data object based on the subset of the dataset and the machine learning algorithm
  6. andimplementing the ML model generation tool in a computer server. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250342404A1 · filed 7/15/2025
• State Farm under examination published 10/30/2025

  Artificial intelligence based systems and methods for document classification and analysis

  Describes extracting, classifying and verifying content from images of insurance claim documents, then applying verified content to the claim or issuing a denial, to automate claims document handling.

  How it works

  seeks to extract content from claim document images, classify the document, verify the content, then apply it to the claim or deny it.

  Where it fits

  would serve claims processing and document intake for insurance.

  Stage

  filed July 2025, under examination, where the claims are commonly narrowed before any allowance.

  What the examiner said one office action

  First action 7/10/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250336002A1 · filed 7/8/2025
• Allstate under examination published 10/23/2025

  System generated damage analysis using scene templates

  Proposes estimating probable causes and value of damage to an object from descriptive and third-party data using machine classifiers, with confidence metrics and scene-diagram interfaces, to support claims damage and liability analysis.

  How it works

  The system draws a geographic scene containing a vehicle built from its operational data and damage indicators plus movable objects; after the user drags objects to new positions it regenerates the scene and produces event interpretation data carrying a confidence metric.

  Where it fits

  Auto claims accident reconstruction and liability determination.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A system, comprising:

  1. at least one processor
  2. memory storing instructions that, when executed by the at least one processor, cause the system to: generate a graphical user interface comprising a graphical representation of a particular geographic location, wherein the graphical representation comprises a graphical representation of a vehicle that is generated based on vehicle operational data associated with the vehicle and at least one indicator of damage to the vehicle, and one or more movable objects
  3. after receiving an indication to move one or more of the one or more movable objects from one or more respective first locations within the graphical representation to one or more respective second locations within the graphical representation, regenerate the graphical user interface
  4. generate, based at least in part on the re-generated graphical representation, event interpretation data, wherein the event interpretation data comprises a confidence metric. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/2/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250329197A1 · filed 6/1/2025
• State Farm awaiting examination published 10/23/2025

  Artificial intelligence-based advanced secure information systems and methods

  Describes storing personal information with per-item privacy settings and, on an access request, comparing the request's attributes to those settings to release only the approved information, as a privacy-governed data access mechanism.

  How it works

  The system stores each individual's personal information with its own privacy settings, receives an access request carrying one or more attributes, compares those attributes against the stored settings, determines which items are approved for disclosure, and transmits just those items to the requestor.

  Where it fits

  Consent-governed personal data access; no insurance-specific workflow is stated.

  Stage

  Filed April 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  1. A computer system for advanced provisioning of secure information, the system comprising at least one processor in communication with at least one memory device, the at least one processor programmed to: store a plurality of sets of personal information for a plurality of individuals, wherein each set of personal information of the plurality of personal information is stored with a plurality of privacy settings for accessing the corresponding set of personal information
  2. receive, from a requestor device, a request for access to a first set of personal information for a first individual, wherein the request for access includes one or more attributes
  3. compare the one or more attributes of the request for access to the plurality of privacy settings for the first set of personal information
  4. determine one or more items of information from the first set of personal information approved to be provided in response to the request for access
  5. generate a response to the request for access to the first set of personal information including the one or more items of information
  6. transmit, to the requestor device, the response to the request for access to the first set of personal information. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250328686A1 · filed 4/3/2025
• TIAA final rejection published 10/23/2025

  Techniques for providing artificial intelligence mediated curation of access and content within a cyber threat intelligence platform

  Proposes a trained AI model that curates cyber threat intelligence objects from a distributed ledger by relevance and recommends a threat practice in response to a queried threat type, to support security operations.

  How it works

  Given a cyber threat type submitted by a node with distributed-ledger access, a trained AI model identifies threat intelligence objects that were contributed by nodes or model-generated, scores each for relevance, and returns a curated set together with a recommended threat-response practice.

  Where it fits

  Cyber security threat intelligence sharing; adjacency to cyber insurance risk is not stated.

  Stage

  Filed April 2024; the claim has already been narrowed by amendment during prosecution, which does not mean the examiner has agreed.

  The claim as amended, showing changes

  A system for providing artificial intelligence (AI) mediated curation of access and content within a cyber threat intelligence platform, the system comprising:

  1. one or more memories storing computer-executable instructions including an Alengine
  2. one or more processors communicatively coupled with the one or more memories that are configured to execute the computer-executable instructions and cause the system to: receive, from a node having access to a distributed ledger, a cyber threat content input indicating a cyber threat type, identify, by a trained Almodel of the Alengine, one or more cyber threat intelligence content objects based on the cyber threat type, each of the one or more cyber threat intelligence content objects being (a) contributed by one or more of one or more nodes having access to the distributed ledger or (b) generated by the trained Almodel, wherein the one or more cyber threat intelligence content objects indicate a response mechanism configured to mitigate effects of the cyber threat type , evaluate, by the trained Almodel, each of the one or more cyber threat intelligence content objects to: determine respective relevance values corresponding to each of the one or more cyber threat intelligence content objects, and generate (i) a curated set of cyber threat intelligence content objects based on the respective relevance values and (ii) a recommended cyber threat practice, and transmit the recommended cyber threat practice and an indication of the curated set of cyber threat intelligence content objects to the node. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 2 office actions

  First action 1/30/2026, most recent 7/14/2026. The grounds raised so far:

  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: US 12,164,664, Lal-Bazelgette-Dupont, Xu et al, Smith et al.

  A further rejection under section 112 appears in this file, in a form this parser reads only from the section heading.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250328634A1 · filed 4/17/2024
• CSAA awaiting examination published 10/16/2025

  Systems and methods for advanced home systems

  Describes generating a home recommendation model from images of a home environment and user input, then surfacing home recommendations on a device, as a property-focused recommendation mechanism.

  How it works

  A mobile device with a camera captures images of a home environment and a related user input; a separate modeling device receives them, builds a home recommendation model with machine learning from the images and input, identifies recommendations, and sends them back to the device to display.

  Where it fits

  Homeowner servicing and property loss-prevention guidance.

  Stage

  Filed April 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  1. A system for generating home recommendations, wherein the system comprises: a mobile computing device, wherein the mobile computing device comprises a camera, a user interface, one or more processors, and a non-transitory computer-readable medium, having stored thereon program instructions that, upon execution by the one or more processors, cause the mobile computing device to perform a set of operations comprising: capturing, via the camera of the mobile computing device, a plurality of images of a home environment
  2. receiving, via the user interface of the mobile computing device, a user input, wherein the user input is associated with the captured plurality of images of the home environment
  3. a modeling computing device, wherein the modeling computing device comprises one or more processors and a non-transitory computer-readable medium, having stored thereon program instructions that, upon execution by the processor, cause the modeling computing device to perform a set of operations comprising: receiving, from the mobile computing device, the captured plurality of images of the home environment and the received user input
  4. generating a home recommendation model using one or more machine learning models, wherein the one or more machine learning models are configured to generate the home recommendation model using the captured plurality of images of the home environment and the received user input
  5. identifying one or more home recommendations, wherein the one or more home recommendations are based on at least the generated home recommendation model
  6. transmitting, to the mobile computing device, instructions that cause the mobile computing device to display, via the user interface of the mobile computing device, a graphical indication of the one or more home recommendations. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250322662A1 · filed 4/4/2025
• State Farm under examination published 10/16/2025

  Assigning data structures to instances of applications using machine learning

  Proposes applying a machine learning model to a policy's data structure and attributes to assign that policy to one of several application instances in a policy administration system, as an infrastructure routing mechanism.

  How it works

  The system identifies a policy's data structure, gathers its attributes, and applies a machine learning model trained on prior policy-to-instance assignments to assign that data structure to one of several application instances of the policy administration system.

  Where it fits

  Back-office policy administration and workload routing for insurers.

  Stage

  Filed April 2024 and in examination; reflects the original claim as filed.

  The claim as filed

  A method of assigning policies across application instances using machine learning (ML) models, comprising:

  1. identifying, by one or more processors of a policy administration system, a first data structure of a first policy
  2. obtaining, by the one or more processors, a first plurality of attributes associated with the first policy
  3. applying, by the one or more processors, a ML model to the first data structure and the first plurality of attributes, wherein the ML model is trained using a plurality of instance assignments, each of the plurality of instance assignments identifying (i) a second data structure of a second policy, (ii) a second plurality of attributes, and (iii) a respective application instance selected from a plurality of application instances based on the second data structure and the second plurality of attributes
  4. assigning, by the one or more processors, from applying the ML model, the first data structure of the first policy to a first application instance from the plurality of application instances of the policy administration system. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250322200A1 · filed 4/10/2024
• Allstate awaiting examination published 10/9/2025

  Processing system having a machine learning engine for providing a surface dimension output

  Describes deriving a surface dimension from a mobile image using standardized reference objects, then computing settlement and repair outputs, to support claims estimation.

  How it works

  The system takes an image, generates bounding boxes resized to a neural network's required dimensions, runs them through the network to find a standardized reference object, derives the real dimensions of a target object from the reference object's pixel size, and transmits those dimensions.

  Where it fits

  Property and auto claims measurement, settlement and repair estimation.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A method comprising:

  1. receiving, by an image analysis and device control system, at least one image
  2. determining, using one or more machine learning algorithms, a plurality of bounding boxes corresponding to the at least one image, wherein determining the plurality of bounding boxes includes adjusting dimensions of the plurality of bounding boxes to match predetermined dimensions for a neural network, and inputting, into the neural network, the plurality of bounding boxes for analysis by the one or more machine learning algorithms to determine whether the at least one image comprises a reference object
  3. determining, by the image analysis and device control system and based at least in part on pixel dimensions of the reference object, actual dimensions of an object comprised within the at least one image
  4. transmitting, by the image analysis and device control system and to another device, the actual dimensions. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250315868A1 · filed 6/16/2025
• State Farm under examination published 10/9/2025

  Reducing false positive fraud alerts for online financial transactions

  Proposes comparing the location of the device used for a card transaction against the authorized cardholder's geolocation and blocking the transaction on a mismatch, to reduce false positive fraud alerts.

  How it works

  The method receives transaction data from the device used for an online purchase and determines its location, obtains a record from a second device tied to the same user to determine where the user actually was, and updates the transaction's status based on whether the two locations agree.

  Where it fits

  Payment fraud detection and reduction of false-positive declines.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A computer-implemented method of detecting fraudulent online transactions based on location data, the method comprising:

  1. receiving, by one or more processors, transaction data associated with an online transaction performed via a first user device associated with a user
  2. determining a first location of the first user device for a time associated with the online transaction
  3. receiving a device record from a second computing device associated with the user
  4. determining a second location of the user at the time associated with the online transaction, based at least in part on the device record
  5. modifying a status of the online transaction, in a digital record, based at least in part on the first location of the first user device and the second location of the user at the time associated with the online transaction. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/15/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Babitch et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250315844A1 · filed 6/17/2025
• Allstate awaiting examination published 10/9/2025

  Systems and methods for reducing false positive error rates using imbalanced data models

  Describes generating multiple resampled imbalanced datasets, building an ensemble of models from them and comparing false positive error rates, to lower the false positive rate of a machine learning process.

  How it works

  Starting from an imbalanced fraud dataset and its baseline false-positive rate, the method resamples it into multiple new datasets each with a distinct positive-to-negative ratio, trains a separate fraud model on each, and ensembles a subset of their outputs to achieve a lower false-positive rate.

  Where it fits

  Fraud detection model development where positive cases are rare.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A method comprising:

  1. receiving, by a processor of a computing device, a first data set having a first imbalanced sampling ratio, wherein the first imbalanced sampling ratio comprises an unequal ratio of positive target data points to negative target data points in the first data set, wherein the first data set comprises information that is indicative of fraud
  2. generating, by the processor, a first model based on the first data set and the first imbalanced sampling ratio, wherein the first model is configured to detect one or more instances of fraud
  3. generating, by the processor, a first output data set based on an input of the first data set to the first model, the first output data set comprising one or more instances of fraud detected by the first model
  4. determining, by the processor, a first false positive error rate for the first output data set, the first false positive error rate comprising a first rate at which fraud is incorrectly detected
  5. resampling, by the processor, the first data set to generate a plurality of new data sets, wherein each new data set of the plurality of new data sets has a respective imbalanced sampling ratio that (i) comprises an unequal ratio of positive target data points to negative target data points in the respective new data set, and (ii) is unique with respect to both (a) the first imbalanced sampling ratio of the first data set, and (b) the imbalanced sampling ratio of each of the other new data sets of the plurality of new data sets
  6. generating, by the processor, a plurality of new models, wherein each new model of the plurality of new models is generated based on a different new data set of the plurality of new data sets and the respective imbalanced sampling ratio of the different new data set, wherein each new model of the plurality of new models is configured to detect one or more instances of fraud
  7. generating, by the processor, a plurality of new output data sets, wherein each new output data set of the plurality of output data sets is generated based on an input of a respective new data set of the plurality of new data sets to a respective new model of the plurality of new models, wherein each new output data set of the plurality of new output data sets comprises one or more additional instances of fraud detected by a respective one of the plurality of new models
  8. determining, by the processor, a plurality of ne

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250315739A1 · filed 6/23/2025
• State Farm under examination published 10/2/2025

  AI/ML chatbot for negotiations

  Describes a chatbot that negotiates terms for one party using a supplied parameter such as priority terms or a negotiating style, judges whether proposed terms are acceptable, and generates counter-terms to send to the other party when they are not.

  How it works

  Retrieves a negotiating parameter set by one contracting party, such as a prioritized product list, a set of priority terms, or a negotiating style, uses the AI chatbot to test whether prospective terms are acceptable against it, and where they are not, generates counter-terms from the parameter and sends them to the other party.

  Where it fits

  Automated negotiation of contractual or policy terms between parties; the abstract does not tie it to a specific insurance line.

  Stage

  Filed June 2025, published October 2025 and under examination with a non-final office action mailed; the claim read here is the original as filed.

  The claim as filed

  A computer system for intelligent negotiation of terms between contracting parties by an artificial intelligence (AI) chatbot, the computer system comprising:

  1. one or more processors
  2. a memory storing executable instructions thereon that, when executed by the one or more processors, cause the one or more processors to: determine or retrieve a parameter indicated by a first contracting party, the parameter including one of: (i) a prioritized product list, (ii) a set of priority contractual terms, or (iii) a negotiating style, determine, via the AI chatbot, whether one or more prospective terms are acceptable terms based on the parameter, responsive to determining that the one or more prospective terms are unacceptable terms, generate, via the AI chatbot, one or more counter terms based on the parameter, and transmit the one or more counter terms to a second contracting party. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/7/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250310284A1 · filed 6/9/2025
• State Farm awaiting examination published 10/2/2025

  Apparatuses, systems, and methods for detecting vehicle occupant actions

  Proposes categorizing in-vehicle image sensor data into driver postures and normalizing them across drivers and camera positions for storage, as a foundation for driver behavior monitoring.

  How it works

  The system receives data from in-vehicle image sensors, categorizes it into driver postures representing driver movements, rotates and scales those postures so they are standardized across different drivers and camera positions, and stores the normalized postures in a database.

  Where it fits

  Telematics and usage-based auto insurance driver-behavior analysis.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A system comprising one or more processors and one or more non-transitory computer-readable media storing computing instructions that, when executed on the one or more processors, cause the one or more processors to perform operations comprising:

  1. receiving sensor data detected by one or more image sensors in a vehicle
  2. categorizing the sensor data as driver postures representative of driver movements in the vehicle
  3. rotating and scaling the driver postures to be standardized for different drivers and for different locations of the one or more image sensors within different vehicles
  4. storing, in a database, the driver postures, as rotated and scaled. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250308263A1 · filed 6/14/2025
• State Farm awaiting examination published 10/2/2025

  Systems and methods for commercial inventory mapping including determining if goods are still available

  Describes using LIDAR and AI to track store inventory, check whether ordered goods remain available and overlay their location on a generated store map, as a retail inventory and mapping mechanism.

  How it works

  From LIDAR data the method generates a virtual map of a store, determines where the still-available goods from an electronic order sit, overlays those locations on a map showing the aisles, and on confirmation that items were picked up or delivered charges a virtual pay account.

  Where it fits

  Retail inventory and order fulfillment; no insurance workflow is stated.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  1. A computer-implemented method for commercial inventory mapping, the computer-implemented method comprising, via one or more local or remote processors, sensors, servers, light detection and ranging (LIDAR) devices, and/or transceivers: generating a LIDAR-based virtual map of a store from processor analysis of LIDAR data
  2. determining locations of goods in an electronic order that are still available
  3. overlaying the determined locations of the goods onto the LIDAR-based virtual map of the store
  4. generating an updated LIDAR-based virtual map of the store displaying aisles of the store and the determined locations of the goods within the store
  5. receiving confirmation that items of a virtual or electronic customer list have been picked up or delivered
  6. charging a virtual pay account for the picked up or delivered items. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250307768A1 · filed 6/13/2025
• Allstate awaiting examination published 10/2/2025

  Vehicle Sharing Optimization

  Proposes using a machine learning algorithm to forecast when vehicle demand will exceed supply in a car-sharing application and prompting users to supply a vehicle, as a fleet balancing mechanism.

  How it works

  The method compiles historical vehicle-sharing supply and demand data, predicts for a given date and location that demand will exceed supply, and on that basis directs a vehicle to drive to a parking location near that place on that date.

  Where it fits

  Vehicle-sharing fleet balancing; no insurance workflow is stated.

  Stage

  Filed June 2025; the original claims were cancelled and this claim substituted, typically after a rejection.

  The claim new claim

  A method comprising:

  1. determining, by a computing device, historical vehicle supply data representing vehicle sharing offers from a plurality of users
  2. determining, by the computing device, historical vehicle demand data representing vehicle sharing requests
  3. based on the historical vehicle supply data and the historical vehicle demand data, determining, by the computing device and for a determined date and a determined location, an expected vehicle demand will exceed an expected vehicle supply
  4. directing, based on the determination that the expected vehicle demand will exceed the expected vehicle supply, a vehicle to drive to a parking location in proximity to the determined location on the determined date. New 3

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250307722A1 · filed 6/11/2025
• Allstate awaiting examination published 10/2/2025

  Neural networks for collision detection

  Describes a combined convolutional and LSTM neural network that scores mobile-device kinematic data over time to predict whether a trigger event is an impact, to support collision detection.

  How it works

  Kinematic variables from a mobile device's movement over a time window are fed to a model whose CNN layer extracts features per time segment, an LSTM layer consumes those in sequence, and a prediction layer produces a score used to classify the event as a type of impact, output with timing.

  Where it fits

  Telematics crash detection and automated first notice of loss for auto insurers.

  Stage

  Filed March 2024 and in examination; reflects the original claim as filed.

  The claim as filed

  A method comprising:

  1. inputting, in a machine-learning collision prediction model, a set of one or more kinematic variables associated with a trigger event based on movement of a mobile device, wherein the set of one or more kinematic variables are recorded over a duration of time, the machine-learning collision prediction model including at least a convolutional neural network (CNN) model layer, a long short-term memory (LSTM) model layer, and a prediction model layer, wherein the CNN model extracts features for time segments of the duration of time, and the respective features associated with respective time segments are fed into the LSTM model layer
  2. generating a prediction score using the machine-learning collision prediction model, the prediction score generated based on the set of one or more kinematic variables and associated with the event
  3. generating a prediction that the event is a type of impact, the prediction generated based on the prediction score
  4. outputting the prediction, wherein the outputted prediction is time-oriented. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250307646A1 · filed 3/26/2024
• State Farm under examination published 10/2/2025

  Systems and methods for analysis of user telematics data using generative AI

  Proposes using a generative AI model to infer a user's personalization characteristics, including information retention, and build a personalized graphical interface, framed around telematics data analysis.

  How it works

  A generative AI model identifies the user from a user action, determines personalization characteristics linked to the user's information retention rate, and generates a personalized graphical interface whose visual graphics are shaped by those characteristics.

  Where it fits

  Personalized customer-facing interfaces; an insurance-specific use is only loosely implied.

  Stage

  Filed June 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A computer-implemented method for analyzing user data and generating apersonalized interface, the computer-implemented method comprising:

  1. determining, by one or more processors, a user identity for a user at a generative artificial intelligence (AI) model based upon a user action
  2. determining, by the one or more processors and based upon at least the user identity, one or more personalization characteristics associated with at least an information retention rate for the user via the generative Almodel, wherein: the information retention rate is indicative of a baseline rate for userunderstanding of information, and the one or more personalization characteristics are predicted to affect the userunderstanding of the information based upon the information retention rate
  3. generating, by the one or more processors, a personalized graphical interface for the user via the generative Almodel, the personalized graphical interface including one or more visual graphics based upon at least the one or more personalization characteristics. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/27/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250307279A1 · filed 6/13/2025
• Elevance under examination published 10/2/2025

  Real-Time and Diagnostic Omnichannel Interaction Insights, Actions, and Management Using Machine Learning Models

  Describes prompting machine learning and large language models with service-channel transcripts and natural language questions over a benefits database to generate omnichannel interaction insights and agent recommendations, for member service.

  How it works

  The method collects transcripts from digital service channels, generates a channel-specific prompt for each from the channel type and extracted metadata, feeds them to machine learning models for insights, and integrates those with member healthcare data using sentiment analysis to display analytical insights.

  Where it fits

  Health plan member service operations and benefits support.

  Stage

  Filed November 2024 and in examination; reflects the original claim as filed.

  The claim as filed

  A method for generating real-time and diagnostic omnichannel interaction insights, the method comprising:

  1. obtaining one or more transcripts corresponding to a plurality of digital service channels
  2. generating one or more channel-specific prompts for the one or more transcripts based on each digital service channel corresponding to a respective transcript and metadata extracted from the one or more transcripts
  3. inputting the one or more channel-specific prompts to one or more machine learning models to obtain channel-specific insights
  4. generating and displaying analytical insights by integrating the channel-specific insights with member-specific healthcare data, using sentiment analysis and data filtering. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/29/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Thondikulam et al, Madisetti et al, Siracusano et al, Dougherty et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250307277A1 · filed 11/6/2024
• Travelers awaiting examination published 10/2/2025

  Digital delegate computer system architecture for improved multi-agent large language model (LLM) implementations

  Describes a multi-agent large language model architecture in which a primary agent checks a user's access entitlements, selects only the tools that user is permitted to use, plans a multi-step response and delegates each step to a secondary agent.

  How it works

  On a user request, an identity server looks up the user's access entitlements; a primary LLM agent uses them to select only the tools that user may access, builds a multi-tier plan assigning each action to a tool, then calls a dedicated secondary LLM agent per tool and assembles their responses into a reply.

  Where it fits

  Governed multi-agent LLM tooling where user permissions constrain which capabilities run; general enterprise AI infrastructure, not an insurance-specific function.

  Stage

  Filed February 2025 and published October 2025; under examination on the claim as originally filed, not yet narrowed in prosecution.

  The claim as filed

  A multi-Large Language Model (LLM), multi-agent, digital delegate computer-implemented method, comprising:

  1. receiving, by a multi-agent LLM server comprising a plurality of electronic processing devices, and from a user device, a user request comprising a prompt and an indication of an identifier of a user of the user device
  2. identifying, by an identity server comprising at least one electronic processing device, the identity server being in communication with the multi-agent LLM server, and utilizing the identifier of the user to query a non-transitory access entitlement data store device in communication with the identity server, the access entitlement data store device storing access entitlement data in relation to user identification data, at least one access entitlement assigned to the user
  3. identifying, by the multi-agent LLM server, and by an execution of instructions defining a primary LLM agent stored in a non-transitory LLM data store device in communication with the multi- agent LLM server, and based on the at least one access entitlement assigned to the user, and by querying data stored in the LLM data store device that is descriptive of a plurality of LLM tools, a subset of LLM tools from the plurality of LLM tools that the user is entitled access to
  4. generating, by the primary LLM agent and utilizing both the prompt and the identified subset of LLM tools from the plurality of LLM tools that the user is entitled access to, and after the identification of the subset of LLM tools from the plurality of LLM tools that the user is entitled access to, a multi-tier plan for responding to the user request, wherein the multi-tier plan defines a plurality of actions, with each action being assigned to one of the LLM tools from the identified subset of LLM tools
  5. executing, by the primary LLM agent, the multi-tier plan, by: (i) calling, based on instructions stored in the LLM data store device defining, for each LLM tool of the plurality of LLM tools, a secondary LLM agent, a secondary LLM agent assigned to each respective one of the LLM tools from the identified subset of tools for the plurality of actions of the multi-tier plan
  6. (ii) receiving, from each secondary LLM agent and in response to the calling, a response for each of the actions of the multi-tier plan
  7. constructing, by the primary LLM agent and utilizing the responses for the actions of the

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250307024A1 · filed 2/24/2025
• Zesty.ai allowed published 10/2/2025

  Hail Predictions Using Artificial Intelligence

  Proposes chaining hail size, hail frequency, damage frequency and damage severity models for a location to estimate hail damage risk, to support property pricing and catastrophe exposure assessment.

  How it works

  For a given location the method trains hail size, hail frequency and damage severity models that can discount older data, determines the location's hail size and frequency, applies a damage frequency model to feature data including those, and applies a damage severity model to property-feature data to estimate severity.

  Where it fits

  Property hail catastrophe pricing, underwriting and loss estimation.

  Stage

  Filed June 2025; the claim has already been narrowed by amendment during prosecution, which does not mean the examiner has agreed.

  The claim as amended, showing changes

  A computer implemented method comprising :

  1. receiving, using one or more processors, a location
  2. training, using the one or more processors, a hail size machine learning model, a hail frequency machine learning model, and a damage severity model, wherein one or more of the hail size machine learning model, the hail frequency machine learning model, and the damage severity model discount older data
  3. determine[[e]]ing, using the one or more processors, a hail size associated with the location using the hail size t hailrsa fimachine learning model
  4. determine[[e]]ing, using the one or more processors, a hail frequency associated with the location using the trsfiahail frequency machine learning model
  5. et of features at the location
  6. rst sociated with the location, and data describing a fissociated with the location, the hail frequency asst feature data including the hail size arsthe fiociated with the location,sst feature data ars, fissorsing the one or more proceusobtain,determine[[e]]ing, using the one or more processors, a damage frequency associated with the location by applying trsa fidamage frequency machine learning model to [[the ]]first feature data including the hail size associated with the location and the hail frequency associated with the location
  7. and et of features at the location
  8. andsthe second feature data including data describing a secondith the location,wociatedss, second feature data assorsing the one or more proceusobtain,determine[[e]]ing, using the one or more processors, a damage severity associated with the location by applying the trsa fidamage severity machine learning model to [[the ]]second feature data including data describing a set of property features at the location. Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250306244A1 · filed 6/16/2025
• Assurant under examination published 9/25/2025

  Systems, methods, and apparatuses for dynamic content extraction in visual media content

  Describes extracting a content object from a selected segment of visual media and generating a relevance object for display, as a general media content extraction mechanism.

  How it works

  The system receives a segment selection for a piece of visual media, identifies the segment from the temporal indicator tied to that selection, extracts a content data object from a portion of the segment, generates a relevance data object from it, and displays that object to the user.

  Where it fits

  General video content extraction; no insurance workflow is stated.

  Stage

  Filed March 2025 and in examination; reflects the original claim as filed.

  The claim as filed

  A system for dynamic content extraction in visual media content, the system comprising one or more processors and at least one non-transitory memory comprising instructions that, with the one or more processors, cause the system to:

  1. receive a segment selection indication associated with visual media content
  2. identify a segment of the visual media content based on temporal indicator associated with the segment selection indication
  3. extract a content data object from at least one portion of the segment of the visual media content
  4. generate a relevance data object based on the content data object
  5. cause display of the relevance data object to a user. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 6/10/2026. The grounds raised so far:

  • section 102, anticipation: the examiner found a single earlier reference showing the same thing
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  References cited against it: Zhang et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250301205A1 · filed 3/21/2025
• Allstate awaiting examination published 9/25/2025

  Processing system having a machine learning engine for providing a surface dimension output

  Proposes measuring an object from a mobile image via standardized reference objects and predicting other objects likely in the room, then estimating a repair cost, to support claims estimation.

  How it works

  The system takes an image, forms bounding boxes matched to a neural network's dimensions, reduces their quality and transposes them onto a standardized image, derives each object's real dimensions from its pixel size, and infers which objects are in the room from those dimensions and objects predicted for that room type.

  Where it fits

  Property claims room contents assessment and repair cost estimation.

  Stage

  Filed January 2025; the original claims were cancelled and this claim substituted in their place, typically after a rejection.

  The claim new claim

  A computing platform, comprising:

  1. at least one processor
  2. a communication interface communicatively coupled to the at least one processor
  3. memory storing computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: receive at least one image
  4. determine a plurality of bounding boxes comprising the at least one image, wherein at least some of the plurality of bounding boxes have dimensions that match predetermined dimensions for a neural network
  5. reduce image quality of the plurality of bounding boxes
  6. transpose the plurality of bounding boxes onto an image having the predetermined dimensions for the neural network
  7. determine respective object dimensions to correspond to respective pixel dimensions of the plurality of bounding boxes
  8. determine, based at least in part on the respective object dimensions and one or more objects predicted by the computing platform to be present in a room type associated with the at least one image, one or more objects that are in the room associated with the at least one image. New 2

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  US20250299231A1 · filed 1/18/2025
• Assurant under examination published 9/25/2025

  Apparatus and method for resource allocation prediction and modeling, and resource acquisition offer generation, adjustment and approval

  Describes predictive models that forecast future resource needs across channels and generate, adjust or approve a resource offer set as demand and value change over time, as a general resource allocation mechanism.

  How it works

  Retrieves offer-generation input data plus a benchmark and portfolio target data set, applies them to an offer-generation model to produce an offer set that meets the targets, renders an adjustment interface for one user to revise it, then routes the revised set to a second user's approval interface and stores it with the resulting status.

  Where it fits

  General resource-offer generation and approval; the benchmark and portfolio targets could touch insurance pricing or underwriting, though the abstract stays domain-neutral.

  Stage

  Filed December 2024 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  A computer-implemented method for generating a resource offer set, the method comprising:

  1. retrieving at least one resource offer generation input data set
  2. receiving a benchmark and portfolio target data set in response to an input by an offer control user via one or more client devices
  3. generating a resource offer set by applying at least one of the at least one resource offer generation input data set and the benchmark and portfolio target data set to a resource offer generation model, wherein the generated resource offer set satisfies the benchmark and portfolio target data set
  4. generating a control signal causing a renderable object comprising an offer adjustment interface displayed at a first of the one or more client devices and configured for updating the resource offer set to create an adjusted resource offer set, the offer adjustment interface comprising an indication of the resource offer set
  5. receiving a completion control signal from the first of the one or more client devices
  6. in response to the completion control signal, generating an approval request control signal causing a second renderable data object comprising an approval interface to be displayed at a second of the one or more client devices, wherein the approval interface comprises an indication of the adjusted resource offer set
  7. receiving, from the second of the one or more client devices, an offer approval control signal comprising an offer status indicator
  8. storing the resource offer set associated with the offer status indicator. SVG 18990898.06-17-2025.MC0V1425X82X102.CLM.1.32.643.3116.653.3142.svg 0.087 0.033 Chemistry Black and white Attorney Docket No: 006128/623995 Reply to Notice to File Missing Parts of January 17, 2025 Original 39

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/15/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step

  Also rejected for double patenting, meaning the examiner reads the claims as too close to another application by the same applicant.

  References cited against it: Barker, Slaight.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250299129A1 · filed 12/20/2024
• Cigna awaiting examination published 9/25/2025

  Machine learning systems for automated database element processing and prediction output generation

  Proposes two machine learning models that predict the likelihood and expected count of avoidable negative health events for each member, then target outreach campaigns accordingly, for population health management.

  How it works

  Trains a model on historical census, lifestyle and employment profile data, then for each person builds a feature vector from that data and computes a score representing the predicted time until they transition to retirement, sorts people into bins by that score, and automatically sends bin-specific campaign material to each group.

  Where it fits

  Would support outreach targeting in life and retirement or health servicing, distributing campaigns by predicted score.

  Stage

  Filed June 2025 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  A computerized method of automatic distributed communication, the method comprising:

  1. training a machine learning model with historical feature vector inputs to generate a retirement score output, wherein: the historical feature vector inputs include historical profile data structures specific to multiple historical entities within a specified age range, and the historical profile data structures include at least one of historical structured lifestyle data, historical structured census, data and historical structured employment data
  2. obtaining a set of entities
  3. for each entity in the set of entities: obtaining at least one of structured census data associated with the entity from a structured census database, structured lifestyle data associated with the entity from a structured lifestyle database, and structured employment data associated with the entity from a structured employment database
  4. generating a feature vector input according to the obtained at least one of the structured census data, the structured lifestyle data, and the structured employment data
  5. processing, by the machine learning model, the feature vector input to generate the retirement score output, wherein the retirement score output is indicative of a predicted time period until the entity transitions to a retirement status
  6. assigning the entity to one of multiple bins according to the retirement score output
  7. for one or more of the multiple bins, automatically distributing structured campaign data associated with the bin to each entity assigned to the bin. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250299108A1 · filed 6/3/2025
• Travelers awaiting examination published 9/25/2025

  Email content extraction

  Describes classifying an incoming email's request type, extracting entities from it and matching them against database records to generate a prepopulated processing request, to automate correspondence intake.

  How it works

  Reads an email at a mail server, runs a classifier to determine its request type, extracts entities from the free-form text when the type is supported, scores the extraction confidence, and above a threshold looks the entities up in a database and builds a new processing request prepopulated from a matching record before deleting the email.

  Where it fits

  Would automate intake and routing of email-based requests such as claims or policy servicing correspondence.

  Stage

  Filed June 2025 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  A computer-implemented method, comprising:

  1. accessing an email message received at an inbox of a mail server
  2. extracting a plurality of correspondence data from the email message
  3. applying a correspondence classifier to the correspondence data to determine a request type of the email message
  4. extracting a plurality of entities from the email message in a free-form format, the extracting performed based on determining that the request type is supported
  5. determining a confidence level of the extracting of the entities
  6. based on determining that the confidence level is above a confidence threshold: performing a lookup of the entities in one or more records of a database
  7. generating a new processing request comprising a plurality of prepopulated data fields populated with the entities based on identifying a match in the one or more records of the database
  8. removing the email message from the inbox of the mail server based on the new processing request. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250298820A1 · filed 6/4/2025
• Allstate under examination published 9/25/2025

  Unauthorized Access Detection

  Proposes computing a vehicle fingerprint from sensor data across operating modes and alerting when later sensor readings deviate from it, as an anomaly-based unauthorized access detection mechanism.

  How it works

  Generates a fingerprint from the vehicle's dynamics sensor data recorded during authorized operation, then for later data computes a difference score between the expected and actual vehicle characteristic, updating the fingerprint when the score stays below a threshold and issuing an alert of potential unauthorized use when it does not.

  Where it fits

  Would serve auto telematics and anti-theft risk assessment for insurers.

  Stage

  Filed April 2025 and published September 2025; in examination, written from a new claim the applicant substituted after the original claims were cancelled.

  The claim new claim

  An apparatus implemented in a vehicle, the apparatus comprising:

  1. at least one sensor configured to measure a characteristic of the vehicle
  2. at least one processor
  3. a communication interface communicatively coupled to the at least one processor
  4. memory storing non-transitory computer-readable instructions that, when executed by the processor, cause the apparatus to: generate a vehicle fingerprint associated with vehicle dynamics data communicated from the at least one sensor during one or more authorized modes of operation of the vehicle
  5. after receiving further vehicle dynamics data communicated from the at least one sensor, determine a difference score representing a difference between an expected vehicle characteristic and an actual vehicle characteristic
  6. after determining that the difference score is below a threshold difference, update the vehicle fingerprint based at least in part on the additional vehicle dynamics data
  7. otherwise, provide an alert to a communication interface indicating a potential unauthorized use or other vehicle issue. New 3

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 6/22/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250296530A1 · filed 4/7/2025
• Guidewire allowed published 9/11/2025

  Inferential analysis using feedback for extracting and combining cyber risk information

  Describes a computer agent that gathers internet-facing information about an entity to assess its cyber security risk, then automatically recommends policy setting changes and network changes to reduce that risk, aimed at cyber underwriting and risk assessment.

  How it works

  Uses an agent to collect Internet-accessible information, infers which of that circumstantial data pertains to a given entity and assigns a certainty level that weights its influence on the risk estimate, then determines a network or security-policy change and sends the resulting recommendation to a client system.

  Where it fits

  Would support cyber insurance risk assessment and underwriting of an organization's exposure.

  Stage

  Filed May 2025 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  A method, comprising:

  1. assessing risk of a cyber security failure in a computer network of an entity, using a computer agent configured to collect information from at least one accessible Internet elements, wherein the assessing of the risk comprises: determining circumstantial or indirect information that is indicative of the entity at least in part by inferring that the collected information relates to the entity based on an inferential analysis
  2. determining a certainty level that the circumstantial or indirect information pertains to the entity, wherein an impact of analysis of the circumstantial or indirect information in assessing the risk of the cyber security failure is determined based on the certainty level
  3. automatically determining, based on the assessed risk, a change to reduce the assessed risk, the change comprising one or more of a computer network change and a cyber security policy criteria setting change
  4. causing a recommendation that is generated based at least in part on the change to be provided over a network to a client system. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250286913A1 · filed 5/21/2025
• State Farm under examination published 9/11/2025

  Systems and methods for evaluating generative artificial intelligence (AI) solutions

  Proposes a tool that prompts a user to describe a proposed generative AI solution, then scores its overall value including the likelihood of a data compromising event, and ranks it against other candidate solutions being considered.

  How it works

  Displays a template collecting the components of a proposed generative AI use case, computes a use score capturing the deployment's overall value together with the likelihood of a data-compromising event, and outputs a priority report comparing that score against other GenAI solutions under consideration.

  Where it fits

  Serves AI governance and risk prioritization; the abstract does not tie it to a specific insurance workflow.

  Stage

  Filed February 2025 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  A computing device for objectively evaluating and prioritizing GEN AI uses cases, the computing device comprising:

  1. at least one processor in communication with the at least one memory, wherein the at least one processor is programmed to: prompt a user to input a plurality of components of a proposed generative artificial intelligence (GEN AI) solution by causing to be displayed on a user computing device a template requesting the plurality of components
  2. in response to receiving the plurality of components, electronically evaluate the proposed GEN AI solution by outputting a use score, wherein the use score represents an overall value of deploying the GEN AI solution including a likelihood of a data compromising event occurring as a result of the deployment
  3. output a priority report including a comparison of the use score for the current proposed GEN AI solution to other GEN AI solutions being considered. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 8/26/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 103, obviousness: the examiner combined earlier references to argue the claim is an obvious step
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  References cited against it: Toledano et al, Goldsteen et al.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250284819A1 · filed 2/5/2025
• State Farm awaiting examination published 9/4/2025

  Parametrizing a quantum gate into a data-encoding variational gate

  Describes converting gates of a quantum feature map into variational data-encoding gates and running machine learning to fit their parameters, a general-purpose quantum machine learning technique with no insurance function stated in the abstract.

  How it works

  Generates a quantum feature map corresponding to a quantum circuit of multiple gates, applies variational data-encoding parameterization to turn at least one gate into a variational data-encoded gate, and runs machine-learning processes over the feature map to determine a value for that gate's variational parameter.

  Where it fits

  The abstract describes a general quantum machine-learning technique and does not specify an insurance workflow.

  Stage

  Filed August 2024 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  A computing device comprising:

  1. at least one processor in communication with the at least one memory, wherein the at least one processor is programmed to: generate a quantum feature map corresponding to a quantum circuit, wherein the quantum feature map includes a plurality of gates
  2. employ variational data-encoding parameterization to the quantum feature map to convert at least one gate of the plurality of gates into a variational data-encoded gate
  3. execute machine learning processes using the quantum feature map to determine a value for a variational data-encoded parameter associated with the at least one variational data-encoded gate. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250278653A1 · filed 8/15/2024
• Experian under examination published 9/4/2025

  Systems and methods for an artificial intelligence/machine learning medical claims platform

  Describes machine learning models that predict denials of medical claims and approvals of resubmitted claims, with automatic clustering that routes claims to workflow queues, aimed at claims processing and resource allocation.

  How it works

  Processes payer remit data to label each historical claim as approved or denied, isolates the denied claims, and applies a resubmission-probability model, trained on historical claim and remit data and validated against an error threshold, to predict each denied claim's likelihood of approval if resubmitted.

  Where it fits

  Would serve health insurance claims processing, denial management and resubmission workflows.

  Stage

  Filed December 2024 and published September 2025; in examination, written from the original claim as filed.

  The claim as filed

  1. A computer-implemented method of deploying a claims resubmission predictive model, the computer-implemented method comprising, as implemented by one or more computing devices configured with specific executable instructions to: access a first set of payer remit data associated with a first set of healthcare claims, a first plurality of patients, a first plurality of provider identifiers, and a first payer entity
  2. process the first set of payer remit data to associate each of a plurality of remit data items with at least one of the first set of healthcare claims and an outcome status indicating either approval or denial for each respective healthcare claim to generate a set of denied claims whose outcome status indicates denial
  3. generate and send a request data package for submission to a server to apply a claims resubmission probability artificial intelligence/machine learning (AI/ML) model which is configured to, for each claim associated with the set of denied claims, predict a likelihood of being approved upon resubmission, wherein the claims resubmission probability AI/ML model has been trained using a first set of historical claim and remit data, and wherein training of the claims resubmission probability AI/ML model comprises: accessing the first set of historical claim and remit data, wherein the first set of historical claim and remit data is associated with a training data set, and wherein the training data set is configured to be updated on a predetermined schedule, accessing, from a data store, one or more model parameters, and based on the first set of historical claim and remit data and the one or more model parameters, training the claims resubmission probability AI/ML model, wherein the training includes validating the claims resubmission probability AI/ML model by determining that an error threshold has been satisfied
  4. receive, from the server, a set of claims resubmission prediction data associated with the set of denied claims and associated with the first set of healthcare claims, the set of claims resubmission prediction data comprising: prediction indicators indicating a likelihood of being approved upon resubmission
  5. access first resubmission parameters associated with a first provider identifier of the first plurality of provider identifiers, the first provider identifier associated with a first subset of the set of denied claims

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250278457A1 · filed 12/18/2024
• State Farm under examination published 8/28/2025

  Initiated automatic claim handling through conversational artificial intelligence (AI)

  Proposes detecting client devices near a weather event, soliciting response data from them by messaging, and estimating how many claim records a locality will generate so that services can be allocated, aimed at claims handling and catastrophe response.

  How it works

  Combines weather data with IoT sensor readings to detect a weather event, identifies nearby users from profile data, activates an app that asks each user event-specific questions, then runs a classification engine over the responses and existing records to estimate the number of new records expected and allocates services accordingly.

  Where it fits

  Would support catastrophe claims intake and resource allocation for insurers after weather events.

  Stage

  Filed February 2024 and published August 2025; in examination, written from a claim the applicant has already narrowed during prosecution.

  The claim as amended, showing changes

  A method for facilitating an electronic record in response to an event, comprising:

  1. receiving, by a server of a service provider, weather data associated with a geographic region, and sensor data from a plurality of Internet of Things (IoT) devices in the geographic region
  2. determining, by the server and based at least in part on the weather data and the sensor data, an occurrence of a weather event in the geographic region
  3. responsive to determining the occurrence of the weather event, and based on profile data associated with a plurality of users, determining, by the server, an identifier of a particular client device associated with a user having an address within a defined proximity of the weather event
  4. causing activation, by the server, of an application on the particular client device, the application generating a user interface on a display of the particular client device, wherein the user interface includes a request for messaging response data associated with the weather event , wherein the request comprises one or morequestions based on the weather event and customized for the user
  5. obtaining, by the server and from the particular client device, the messaging response data
  6. determining, by the server and based on the messaging response data, whether the particular client device will generate an electronic record associated with the weather event
  7. applying, by the server, an algorithma classification engineto comparethe messaging response data from the particular client device [[to]]and electronic records associated with the service provider to output[[
  8. ]] based on comparing the messaging response data to the electronic records, determining, by applying the algorithm,[[a]]an estimated quantity of new electronic records to be generated at a locality in the defined proximity of the weather event
  9. based at least on the estimated quantity, determining, by the server, at least one service from a plurality of services to be allocated to assist in facilitating a processing of the new electronic records generated at the locality. SVG 18585902.05-19-2026.MPD0CNGZ4X72X80.CLM.1.30.1900.3024.2250.3127.svg 0.343 1.167 Chemistry Black and white Currently amended 2

  Underlined matter was added during prosecution and struck matter was removed, as 37 CFR 1.121 requires the applicant to show. An amendment means the applicant moved, not that the examiner agreed.

  Verbatim from the applicant's amended claim at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said 3 office actions

  First action 6/18/2025, most recent 7/1/2026. The grounds raised so far:

  • section 101, eligibility: the examiner questions whether this is patentable subject matter at all
  • section 112, definiteness or written description: the examiner says the claim is unclear or unsupported by the specification

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250272756A1 · filed 2/23/2024
• State Farm under examination published 8/28/2025

  Electronic management of license data

  Describes querying a database of user licenses, detecting which are in a renewal period, notifying those users, and pre-populating a renewal application for their approval, aimed at agent and adjuster licensing administration.

  How it works

  Feeds a candidate user's profile into a model trained on sample roles, practice locations and license data, infers the person's role or location of practice, compares their licenses against those held by similar profiles, and flags a potential licensure discrepancy when a peer-held license is missing before notifying the user.

  Where it fits

  Would support agent and producer licensing compliance and renewal management.

  Stage

  Filed May 2025 and published August 2025; in examination, written from the original claim as filed.

  The claim as filed

  A computer system for identifying licensure discrepancies, the computer system comprising at least one processor in communication with at least one memory, wherein the at least one processor is configured to:

  1. retrieve a candidate user profile from the at least one memory, wherein a plurality of user profiles, including the candidate user profile, are stored in the at least one memory
  2. input candidate user data associated with the candidate user profile to a machine learning model, wherein the machine learning model is trained based upon sample user profiles comprising at least one of sample roles, sample locations of practice, or sample license data, and wherein the machine learning model is configured to, based upon the candidate user data: determine at least one of a role of the candidate user profile or a location of practice of the candidate user profile based upon the candidate user data
  3. identify license data associated with user profiles, of the plurality of user profiles, associated with at least one of the role or the location of practice
  4. compare at least part of the candidate user data to the license data
  5. identify a potential licensure discrepancy for the candidate user profile based upon the comparison, the potential licensure discrepancy associated with a license included in the license data and not included in the candidate user data
  6. receive an output from the machine learning model identifying the potential licensure discrepancy
  7. transmit a notification associated with the potential licensure discrepancy to a user computing device associated with the candidate user profile. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  What the examiner said one office action

  First action 7/6/2026. The only rejection so far is for double patenting.

  Read from the examiner's own office actions at the USPTO. A rejection is a stage in prosecution, not a final answer: most applications are rejected at least once and many are granted anyway.

  US20250272699A1 · filed 5/13/2025
• State Farm awaiting examination published 8/28/2025

  Systems and Methods to Leverage Unused Compute Resource for Machine Learning Tasks

  Describes identifying computing nodes in an inactive state and scheduling machine learning sub-tasks to run on them, then collating the results, a general-purpose resource-utilisation method with no insurance function stated in the abstract.

  How it works

  Uses a resource model to estimate a sub-task's resource requirement, schedules that sub-task to run on a computing node during the node's inactive state, collates the finished sub-task into the completed task, and generates a notification when the task is done.

  Where it fits

  The abstract describes general compute-resource scheduling and does not specify an insurance workflow.

  Stage

  Filed May 2025 and published August 2025; in examination, written from the original claim as filed.

  The claim as filed

  A system for leveraging inactive computing resources to optimize task scheduling and performance among available computing resources, comprising:

  1. one or more computing nodes having an inactive state indicating deactivation
  2. one or more processors
  3. a memory communicatively coupled to the one or more computing nodes and the one or more processors, the memory containing instructions therein that, when executed, cause the one or more processors to: determine, based upon a resource model, a resource requirement for a sub-task of a task for execution on the one or more computing nodes, schedule, based on the resource requirement, the sub-task for execution on a computing node of the one or more computing nodes during the inactive state of the computing node, collate the sub-task into a completed task, and generate a completed task notification indicating the completed task. 2

  Verbatim from the applicant's claim as filed at the USPTO. A pending claim is a request, not a granted right.

  US20250272160A1 · filed 5/8/2025