Allstate's Patent Puts a GPT Between the FNOL Call and the Payout Model

Allstate was granted US Patent 12,705,674 B2 on August 11, 2026 for a claims pipeline in which a generative model reads a policyholder's free-text account of an accident, extracts more than twenty structured claim variables from it, and keeps generating follow-up questions until the required variable set is complete (USPTO Official Gazette, August 11, 2026). Those variables then go to a model that correlates them to claim outcome, including payout amount.

The grant is one line in a portfolio that already runs to 136 insurer AI patents at Allstate alone, part of a three-carrier bloc holding 77% of all insurer AI filings since 2014 (Evident Insurance AI Patent Tracker, via Insurance Journal, December 2025). What separates this one from the intake-chatbot filings around it is where the language model sits in the chain. It is not answering the policyholder. It is manufacturing the covariates that a downstream severity model consumes, which puts a stochastic text parser upstream of a number that eventually reaches the case reserve.

Patent Details

Patent number	US 12,705,674 B2
Title	Claims processing with extracted claim variables
Assignee	Allstate Insurance Company, Northbrook, IL
Inventors	Olivia Mahler-Haug, Corey Fyock, Isabella DeRosa, Nicole Bauer, Julio Castaneda, Daniel Kashkett
Application	18/456,336, filed August 25, 2023
Prior publication	US 2025/0069148 A1, February 27, 2025
Granted	August 11, 2026 (Official Gazette Vol. 1549, No. 2)
Classification	CPC G06Q 40/08, insurance data processing
Claims	19 (system, method, and computing-system sets)

What the Model Is Told to Extract

The specification does not describe a general-purpose summarizer. It names a predetermined required set of claim variables the orchestration circuit has to fill before a claim can move, and the list reads like an adjuster's opening script rendered as a schema:

Whether a tow removal was involved, and the insured vehicle's body type
Whether the vehicle's doors remained operable after the incident
Whether fluid leaks were present, and a gross damage assessment
Whether personal injury occurred, plus injury detail
Whether a police report was filed, and whether prior damage existed
Location down to country, state and city, plus date, time and year
Whether the loss occurred in a no-fault state, and whether texting or phone restrictions apply
Policy coverage status and liability determination

Group those and three distinct actuarial families fall out. Tow, door operability, fluid leaks, body type and gross damage are physical-damage severity proxies, the variables that separate a drivable fender claim from a probable total loss. Injury occurrence and injury detail open the bodily-injury coverage, where severity distributions are heavier and development is longer. The jurisdictional variables, no-fault status and distracted-driving statute applicability, are liability-determination inputs, and their presence is what marks this as something other than an intake convenience. A system extracting no-fault status is not collecting contact details. It is assembling the record on which coverage and liability get decided.

The generative component is explicit in the specification, which describes using "a generative artificial intelligence (AI), such as a Generative Pre-trained Transformer (GPT)" further trained on insurance claim statements (US 12,705,674 B2, USPTO, August 2026). The background section is equally direct about what it displaces, describing conventional claim intake as reliant on "scripted telephone calls" and human-to-human interaction between insureds and adjusters.

The Completeness Loop

Claim 1 recites the loop rather than the extraction. The system receives a first natural-language statement, parses it, compares the extracted variables against the required set, identifies which are missing, generates a request for a second statement targeting those specific gaps, and transmits it. Dependent claims 12, 16 and 17 add the interface: a graphical feature displaying a question about the missing variable, and a second graphical feature offering multiple-choice input for it. Claim 19 closes the loop by requiring an indication that the combined variable set is complete before everything is handed to the claims processing server.

That loop solves a data problem actuaries usually meet on the back end, as a nuisance in a modeling dataset. Missing fields in historical FNOL data are almost never missing at random. Adjusters ask more probing questions when a claim smells expensive, so injury detail and prior-damage history tend to be populated on exactly the files where severity runs high, and blank on the routine ones. Any early-severity model fit on that data is learning partly from the fact of a field being answered rather than from its value. A system that enforces the same required variable set on every claim changes the missingness mechanism itself, and that is a covariate shift, not a data-quality improvement. A severity model validated on pre-adoption FNOL records does not automatically hold on post-adoption ones, because the population of populated fields is no longer the same population.

From Free Text to Payout Estimate

Claims 5, 6, 11 and 13 carry the piece that matters for reserving. The extracted variables go to a claims processing server, the system receives a claim outcome back, and that outcome is displayed to the user device. The specification says what generates the outcome: a statistical, mathematical or machine learning model that correlates the extracted variables to payout amount, vehicle damage severity, total loss determination and at-fault party identification.

Read the architecture carefully and the language model is not the severity model. The GPT produces regressors; a separate correlation model consumes them. That split matters because it splits validation responsibility in a way most model risk frameworks are not set up to handle. Extraction accuracy is a natural-language problem owned by a data science team and measured in field-level precision and recall. The outcome model is a statistical problem owned, in a well-run shop, by actuaries. Neither team owns the interaction.

The interaction is the classic errors-in-variables problem, and it does not behave intuitively. Measurement error in an explanatory variable attenuates that variable's fitted coefficient and pushes the lost signal into the residual. So a correlation model calibrated on hand-coded historical variables, then deployed on GPT-extracted ones, will systematically under-respond to precisely the variable the extractor reads least reliably. If door operability is extracted correctly 96% of the time and total-loss propensity leans heavily on it, the deployed model does not fail loudly on 4% of claims. It runs slightly flat on all of them, and the residual variance it exports looks like ordinary process noise rather than a calibration defect. Field-level extraction accuracy, reported on its own, will not reveal that. What reveals it is a back-test of predicted against realized payout, stratified by whether the variable was machine-extracted or human-entered, which is a test the actuarial side has to ask for.

Where This Touches the Reserve

Case reserves and the development pattern

Allstate carried $41.1 billion of reserves for claims and claims adjustment expense at December 31, 2025, down from $41.9 billion a year earlier (Allstate Form 10-K, filed February 20, 2026). Auto claims are the bulk of the count, and initial case reserves on those files are set early, off exactly the sort of intake information this patent structures.

If a model-derived severity estimate seeds or influences the opening case reserve, the level of case reserve adequacy at early maturities moves. That is not an abstract worry; it is the specific condition the standard reserving correction was written for. Berquist and Sherman's adjustment restates historical case reserves to the current adequacy level before development factors are selected, precisely because incurred development factors drawn from years with different case reserve strength do not describe the current year (Berquist and Sherman, Casualty Actuarial Society, 1977).

The difficulty is that the adjustment presumes the change can be identified and measured. A claims department that revises its case reserving standards issues a memo with a date on it, and the reserving actuary can cut the triangle around that date. An extraction model that retrains on synthetic statements and drifts toward slightly different variable values quarter by quarter produces a continuous, undocumented change in the same quantity. There is no memo, no effective date, and no step change in the diagonal to catch the eye. The adequacy shift shows up as a mild, persistent bias in incurred development that a standard chain-ladder selection absorbs into the factors and carries forward.

The synthetic-data loop

Claims 3, 4, 9, 10 and 15 add a synthetic statement generating circuit that produces new claim statements by modifying originals, with the extraction models then trained on the synthetic set. The specification's stated rationale is coverage of scenarios where, in its words, "claim scenario data will not be frequent enough" for effective training.

Rare scenarios in a personal auto book are where the severity tail lives. Augmenting training data there is a defensible engineering choice and a genuinely awkward actuarial one, for two reasons. First, a generator built from realized outcomes encodes the historical outcome distribution into the extractor's priors, and if generation is conditioned on realized payouts in any form, the extractor is partly learning to anticipate the answer rather than to read the text. Held-out validation on synthetic data will look excellent while performance on a genuinely novel loss type degrades, which is target leakage wearing a data-augmentation costume. Second, and more subtly, the synthetic tail is a resampling of the tail already observed. The model becomes most confident in exactly the region where the book has the least real experience, and that confidence propagates into the payout correlation as narrower predicted variance on tail claims.

For a reserving actuary the operational consequence is short: a stable extraction-accuracy metric is not evidence of stable severity estimation on emerging loss types. Confidence and accuracy are not the same axis once a share of the training corpus is machine-generated, and only realized-outcome back-testing on real claims distinguishes them.

The LAE Arithmetic Across 26 Million Auto Policies

Scale is what turns this from a claims-technology filing into an expense-line item. Allstate reported 25.95 million auto policies in force and $14.9 billion of Property-Liability premiums earned in the second quarter of 2026, with an auto insurance combined ratio of 83.3, improved 2.7 points from the prior year quarter, and an underlying auto combined ratio of 87.6, improved 0.2 points (Allstate, August 6, 2026). Property-Liability underwriting income for the quarter was $2.0 billion against $1.3 billion a year earlier. "Allstate delivered strong operating and financial results in the second quarter of 2026, while executing our strategic growth plans," chief executive Tom Wilson said in the release (Allstate, August 6, 2026).

Now look at the variable list again as a routing schema rather than a severity schema. Tow involvement, door operability, fluid leaks, gross damage and police-report status are the exact fields that decide whether a vehicle goes to a shop, gets a photo-based desk appraisal, or is flagged as a probable total loss. Capturing them at first contact, without the scripted call the background section describes, removes handling steps before an adjuster ever opens the file. That is allocated loss adjustment expense on the routing side and unallocated loss adjustment expense on the staffing side, both compressed by the same mechanism.

Compression of that kind creates its own reserving question, which the site examined in Lemonade's 5% LAE ratio against a roughly 9% incumbent norm. The two workhorse ULAE methods, paid-to-paid and the claim-count-based Johnson method, both assume a stable relationship between claims department cost and either paid loss or claim count. A pipeline that changes how much human handling a standard claim consumes breaks that stability mid-cycle, and it does so gradually rather than at a datable boundary, which is the same identification problem the case reserve drift creates one section up.

Where the NAIC Bulletin Lands

Nothing about this architecture sits outside existing regulatory scope, and the model bulletin's own language is what puts it inside. The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted by Executive Committee and Plenary on December 4, 2023, states that an AI Systems Program "should address the use of AI Systems across the insurance life cycle, including areas such as product development and design, marketing, use, underwriting, rating and pricing, case management, claim administration and payment, and fraud detection" (NAIC Model Bulletin, December 2023). Claim administration and payment is named. Roughly two dozen states and the District of Columbia have issued it, per the NAIC's own adoption map.

Two provisions bite harder here than the general governance language. The bulletin expects the program to cover all phases of an AI system's life cycle, naming "on-going monitoring, updating and retirement" among them, and a synthetic-retraining circuit that revises the extractor against realized outcomes is an updating event on a recurring schedule rather than a one-time deployment. The bulletin also expects the governance structure to draw on committees that include actuarial alongside claims, compliance and legal, which is the formal hook for the split-ownership problem described earlier. Behind both sits the Unfair Claims Settlement Practices Act, Model #900, which the bulletin cites directly and which the bulletin says applies regardless of the methods an insurer used to determine or support its actions.

Three Carriers, Three Segments of the Same Pipeline

Set this grant next to the other claims-side patents issued to large personal-lines carriers this year and a division of the claim lifecycle becomes visible. USAA's storm-boundary and customer-density severity code, granted July 14, 2026, operates before any loss report exists. Allstate's grant operates at the moment of first notice, converting the policyholder's own words into structured variables. Travelers' bodily-injury patent reads medical files well after the claim is open. Three carriers, three non-overlapping segments of the same pipeline, each staking IP on a different point at which a model rather than an adjuster produces the severity signal.

The eligibility angle is worth a line as well. G06Q 40/08 is the classification where insurance software claims were routinely rejected under the Alice framework for most of the last decade. A granted claim set reciting machine-learning extraction plus conditional generation of information requests is a further data point on where examiners are now drawing the abstract-idea line, a shift the site tracked in its coverage of the Section 101 reset in insurance AI patents.

Questions the Reserving Side Can Put to the Claims Side

Does any output of the correlation model, directly or through a recommended range, influence the initial case reserve? If it does, the effective date of deployment becomes a triangle boundary.
Is realized payout ever an input to the synthetic statement generator, in any conditioning form? A yes makes held-out extraction accuracy uninformative about tail performance.
How often does the extraction model retrain, and is each retrain versioned and dated in a way that can be joined to accident period?
Are predicted-versus-realized payout back-tests stratified by machine-extracted against human-entered variables, so attenuation from extraction error is separable from ordinary model error?
Does the ULAE claim-count or paid-loss base still measure the same thing after intake handling changes, or does the historical ratio need refitting before the next reserve review?

None of those questions has a technical answer that a claims organization cannot produce. What they need is for someone to ask before the first post-adoption accident year closes rather than after, because every one of them is far cheaper to answer prospectively than to reconstruct out of a diagonal that has already drifted.