EXL’s 10 AI Patents: How a $2 Billion Services Company Is Building Insurance’s AI Infrastructure

In February 2026, EXL (NASDAQ: EXLS) announced that it had been granted 10 new U.S. patents in a single year for AI and data innovations spanning insurance, healthcare, retail, utilities, and financial services. The announcement came just weeks before EXL reported $2.09 billion in 2025 revenue, with insurance representing its largest segment at $710.6 million.

For actuaries tracking how AI intellectual property is reshaping insurance operations, EXL's patent portfolio introduces a third distinct IP strategy to the competitive landscape we have been mapping since AIG's agentic underwriting patents surfaced in early 2025. Where AIG patents protect a proprietary system built for its own underwriting operations and Quantiphi's Dociphi patents protect a standalone vendor platform, EXL's patents protect methods that a services company deploys on behalf of carrier clients within long-term engagement contracts. The implications for carriers evaluating build, buy, or outsource decisions are significant, and the questions around IP ownership of processed outputs are ones that actuarial leaders and CIOs will increasingly need to confront.

This article maps EXL's complete patent portfolio, examines what each patent actually protects at the claims level, identifies the natural groupings across the 10 patents, and positions EXL's IP strategy against AIG and Quantiphi in what is becoming a three-way patent race for AI dominance in insurance operations.

EXL: The Services Company With $710 Million in Insurance Revenue

Understanding EXL's patents requires understanding EXL's business model. EXL is not an insurance carrier. It does not underwrite policies, hold reserves, or file statutory statements. EXL is also not a pure-play technology vendor selling packaged software. Instead, EXL operates as a data and AI services company that embeds its capabilities within long-term client engagements, processing carrier data, managing carrier operations, and deploying AI tools within carrier workflows.

From tracking EXL's financial disclosures over the past two years, several data points stand out. EXL generated $2.09 billion in total revenue in 2025, representing 13.6% year-over-year growth. Its insurance segment alone reached $710.6 million, making it the company's largest business line at roughly 34% of total revenue. EXL employs approximately 63,000 people across six continents and has guided 2026 revenue to $2.275 to $2.315 billion, implying continued double-digit growth. The company also authorized a $500 million share repurchase program in February 2026, signaling confidence in its cash generation trajectory.

Within insurance, EXL provides services spanning claims processing, underwriting support, subrogation (where EXL claims to be the largest provider in the United States through its Subrosource platform), regulatory reporting, and customer operations. This operational depth is important context for understanding the patents: EXL is not patenting theoretical AI capabilities. It is patenting methods that its 63,000 employees use daily to process carrier data at scale.

The Complete Patent Portfolio: 10 Patents in Three Clusters

From our analysis of all 10 patent filings, the portfolio naturally clusters into three groups based on what each patent protects and how it maps to EXL's disclosed product capabilities.

Cluster 1: Data Ingestion and Knowledge Structuring (3 patents)

These three patents form a coherent pipeline: extract data from unstructured documents, recognize entities within that data, and structure the results into a searchable knowledge graph.

Patent US 12,260,342 (granted March 25, 2025) covers EXL's Xtrakto.AI platform and protects methods for multimodal table extraction and semantic search within unstructured documents. The independent claims describe a system that receives a document with unstructured data, uses a trained neural network to extract alphanumeric data by generating bounding boxes for detected cells, identifies globally applicable items (such as document headers or patient identifiers) from images within the document, generates a searchable data structure, and then performs semantic search against that structure to answer natural-language queries. The patent specifically claims CNN-based cell detection in tables, graph neural network-based entity extraction, and the generation of coordinate-referenced searchable outputs. This is EXL's foundational document ingestion patent, directly comparable to AIG's Patent 12,437,155 (the markdown-based table and text extraction patent), though EXL's approach uses bounding box detection and semantic search rather than AIG's chunking and ontological data store approach.

Patent US 12,353,832 (granted July 8, 2025) covers EXL's Generic NER product and protects a method for context-based named entity recognition that goes beyond conventional NER by attaching domain-specific context to extracted entities. The independent claims describe an AI-based entity extraction and labeling process, followed by a novel "reverse question-and-answer" technique where the system generates questions about extracted entities (for example, "What is value '1978454970' of type 'numeric'?") to predict entity keys like "order id." Those predicted keys are then aligned to subscriber-specific ontologies using similarity matching (Levenshtein distance, Jaro-Winkler distance, or Longest Common Subsequence). The patent specifically addresses the problem of deploying NER across multiple domains without retraining, since the entity alignment step adapts generic extraction to client-specific data schemas at runtime.

Patent US 12,482,215 (granted November 25, 2025) covers EXL's knowledge graph creation capabilities and protects methods for dynamically segmenting documents into content chunks, extracting concepts and relationships, generating graph-based node triples (subject-predicate-object structures), creating vector embeddings indexed to those nodes, and querying the embeddings for semantic search. The claims specifically address the context window problem that limits direct LLM processing of large document sets: rather than feeding entire documents into an LLM, the system constructs a knowledge graph that captures relationships across documents and can be queried efficiently. A key differentiator is the patent's incremental update mechanism, where changes to source documents trigger targeted updates to affected nodes and embeddings rather than requiring full graph regeneration. The patent also claims pre-processing modules for non-textual data, including computer code analysis (via static and dynamic analysis), audio transcription, and computer vision-based image interpretation.

Together, these three patents describe a pipeline: Xtrakto.AI extracts structured data from unstructured documents; Generic NER identifies and contextualizes entities within that data; and the knowledge graph patent structures everything into a queryable, incrementally updatable graph. This pipeline is applicable across all of EXL's industry verticals, not just insurance.

Cluster 2: Insurance-Specific Domain AI (3 patents)

These patents directly serve insurance workflows and represent EXL's most strategically significant IP from an actuarial perspective.

Patent US 12,399,924 (granted August 26, 2025) is the flagship patent covering EXL's Insurance LLM, described in the claims as a "multi-domain signal evaluation system." Despite the abstract terminology, the patent's detailed description is explicitly about insurance. The system receives unstructured alphanumeric signal data (insurance documents containing claims narratives, medical records, legal frameworks) associated with a "decision logic condition" (an insurance claim scenario). It identifies non-compliant signals (PII/PHI information), generates masking elements that preserve contextual metadata while anonymizing protected information (using SHA-256 hashing that produces consistent values for the same entity across different document sections), assigns signal domain categories (medical, legal, financial), generates a composite narrative summary, and produces guidance artifacts with recommended actions and predicted outcomes for specialized users including, as the patent specifically names, "claim adjusters, nurses, underwriters, case managers, assistants, recovery specialists, auditors and adjudicators, actuaries, clinical teams, and contract negotiators."

The patent describes an ensemble architecture where multiple expert models are each trained on different aspects of insurance (claims processing, fraud detection, risk assessment, customer support), with a gating network that dynamically selects the most relevant expert for each task. The training pipeline uses LoRA (Low-Rank Adaptation) for fine-tuning, tensor parallelism for GPU computation, and a customized RAG pipeline. EXL claims the system achieves 20 to 30% higher accuracy on insurance tasks than general-purpose models like GPT-4, Claude, and Gemini, a claim supported by the NVIDIA partnership announced in September 2024 that provided the NeMo framework, H100 GPUs, and Triton Inference Server infrastructure.

This is the most insurance-specific patent in EXL's portfolio. Unlike AIG's patents, which focus on document extraction for underwriting submission intake, EXL's Insurance LLM patent covers the analytical layer: taking already-extracted data and generating actionable recommendations for claims adjudication, underwriting decisions, and risk assessment.

Patent US 12,468,696 (granted November 11, 2025) covers EXL's Regulatory Reporting Assist.AI product and protects a signal evaluation platform for processing regulatory reporting data. The independent claims describe a system that receives natural-language queries about digital artifacts (financial statements, regulatory filings) of a monitored system (such as a statutory insurance reporting cycle), retrieves comparative artifact sets from different time periods, calculates a performance differential report (flux analysis) identifying changes in operational performance characteristics, determines relevant historical dialogue records from prior regulatory queries and responses, and uses a generative ML model to create narrative responses based on the differential report and historical context.

For actuaries, this patent is particularly relevant because the description explicitly references statutory insurance reporting, NAIC compliance requirements, validation of financial statement line items, flux analysis for quarter-over-quarter and year-over-year comparisons, and anomaly detection in key financial ratios and capital adequacy metrics. The system maintains a query repository of prior regulatory inquiries and accepted responses, enabling it to learn from historical regulator interactions when generating new explanations. The patent also claims automated validation against compliance schemas, error resolution using historical correction records, and a materiality indicator that prioritizes validation checks based on regulatory significance and financial impact.

This patent fills a gap that neither AIG nor Quantiphi has addressed. AIG's patents are focused on underwriting document processing. Quantiphi's patents target loss run extraction and submission comparison. EXL's regulatory reporting patent targets the back-office statutory compliance workflow, a domain where actuarial sign-off is mandatory and where automation has been historically limited to spreadsheet-based tools.

Patent US 12,387,271 (granted August 12, 2025) covers EXL's Property Insights product and protects a system for generating event predictions (risk scores, severity estimates, frequency projections) using cognitive image analysis of aerial property imagery. The independent claims describe a multi-model pipeline: a first ML model extracts image-based attributes (roof type, building materials, vegetation density, number of chimneys), a second ML model determines distances between objects of interest (property to vegetation, property to fire station, property to water bodies, window to chimney), and a third ML model generates event predictions by combining image attributes, distance measurements, and location-based attributes (historical claims data, policy information, weather history). The system reduces network traffic by querying local data stores before remote aerial imagery providers and includes mechanisms for comparing properties to sets of similar locations using Euclidean distance-based similarity scoring.

From a property and casualty actuarial perspective, this patent targets the inspection and underwriting workflow for homeowners and commercial property lines. The system generates risk, severity, and frequency predictions for environmental events (fire, flood, storm, earthquake, tornado) and presents them through interactive dashboards showing geographic distributions, attribute-claim correlations, and portfolio-level risk summaries. This directly overlaps with capabilities offered by vendors like Cape Analytics and Nearmap, though EXL's patent claims are specifically structured around the multi-model architecture and network traffic optimization rather than the imagery analysis alone.

Cluster 3: Cross-Industry Enterprise AI (4 patents)

The remaining four patents protect general-purpose AI capabilities that EXL applies across insurance and other verticals. These are individually useful but do not form a coherent insurance-specific system:

US 12,400,252 (EXL Transaction Insights) covers AI-based consumer financial risk mining, including income stability analysis and debt obligation assessment. US 12,334,077 (EXL Smart Audit.AI) covers automated audio-to-text signal processing for quality audit. US 12,299,427 (EXL SAMAC Copilot) covers automatic calibration of AI queries in a structured manner. US 12,536,551 (EXL Paymentor) covers tensor-based reinforcement learning for optimizing multichannel customer communication timing, tone, and content.

These patents are relevant to insurance operations (Transaction Insights for underwriting risk assessment, Smart Audit for claims call quality, Paymentor for collections optimization), but their claims are written broadly enough to apply across banking, healthcare, retail, and other EXL verticals.

The Three-Way Patent Race: Carrier vs. Vendor vs. Services Company

With EXL's portfolio now mapped alongside AIG's and Quantiphi's, a clear pattern emerges in how different types of companies are approaching AI intellectual property in insurance.

AIG (the carrier) holds three patents protecting its proprietary underwriting submission intake pipeline. AIG's patents cover the specific methods by which it extracts data from E&S submissions (markdown-based table and text separation), ensures traceability and detects hallucinations in LLM outputs (chunk-level identification with source attribution), and processes complex multi-table spreadsheets (chain-of-thought prompting). AIG built these systems for its own 370,000+ annual E&S submissions using a three-layer stack of Anthropic Claude, Palantir Foundry, and proprietary tools. AIG's IP strategy is defensive: it protects the specific methods AIG uses so competitors cannot replicate the exact architecture, but the patents are narrow enough that carriers building different extraction approaches can likely design around them.

Quantiphi/Dociphi (the vendor) holds three patents protecting its standalone document processing platform. Quantiphi's patents cover loss run extraction, a three-way comparison screen with human-in-the-loop validation, and enhanced document search capabilities. Dociphi is sold as a platform that carriers and brokers license and deploy. Quantiphi's IP strategy is product-protective: it secures the specific features that differentiate Dociphi from competing platforms like Indico Data, Roots Automation, and other IDP vendors.

EXL (the services company) now holds 10 patents protecting methods used across a much broader operational surface area. Unlike AIG, which patents for internal use, or Quantiphi, which patents for platform sales, EXL patents methods that it deploys on behalf of carrier clients within managed service engagements. This creates a fundamentally different IP dynamic.

When a carrier contracts with EXL for claims processing and EXL uses its patented Insurance LLM to generate adjudication recommendations, several questions arise that do not exist in the AIG or Quantiphi models. Does the carrier have any IP exposure from EXL's patented methods being applied to the carrier's data? Can the carrier replicate the methods internally if the engagement ends? If the carrier switches to a competitor services provider, are there process dependencies on EXL's patented workflows?

These are not hypothetical concerns. EXL's insurance segment generated $710.6 million in 2025 revenue, suggesting deep, multi-year engagements with major carriers. The patents create structural switching costs that go beyond typical service contract dependencies, because the analytical methods themselves are now proprietary.

What EXL Has NOT Patented

In our analysis of AIG's patent portfolio, we identified several capabilities that AIG chose not to patent, which represented potential freedom-to-operate opportunities for other carriers. A similar analysis of EXL's portfolio reveals notable gaps.

EXL has not patented its specific training data architecture or the proprietary datasets used to fine-tune the Insurance LLM. The patent describes the training pipeline (LoRA, supervised fine-tuning, NVIDIA NeMo framework) but these are standard techniques. The competitive moat around the Insurance LLM is the 10+ years of domain-specific labeled data, not the training methodology itself.

EXL has not patented specific claims adjudication decision trees or underwriting rules. The Insurance LLM patent protects the multi-domain signal evaluation architecture, not the specific business logic that determines whether a claim should be approved or denied.

EXL has not patented methods for actuarial reserving, pricing, or capital modeling. Despite naming actuaries as target users in the Insurance LLM patent, none of the 10 patents address core actuarial analytical workflows like loss development, credibility weighting, or stochastic modeling.

EXL has not patented its federated learning architecture in a standalone patent, despite describing it extensively in the Insurance LLM patent as a mechanism for multi-institutional collaboration without sharing sensitive data. This capability is described but not independently claimed.

A Note on the Press Release: Two Patent Number Errors

In the course of verifying each of EXL's 10 patents against the actual USPTO filings, we identified two incorrect patent numbers in EXL's February 9, 2026 press release. The press release lists US 12,481,215 for the knowledge graph patent, but the correct number is US 12,482,215. Similarly, the press release lists US 12,253,832 for the Generic NER patent, but the correct number is US 12,353,832. Both errors appear to be digit transpositions. The correct patents match EXL as the assignee, match the stated grant dates, and match the described capabilities. We note this for readers who may attempt to look up the patents using the press release numbers.

Actuarial Implications

EXL's patent portfolio signals several developments that actuarial professionals should be tracking.

First, the regulatory reporting patent (US 12,468,696) directly targets workflows where actuarial sign-off is required. Automated flux analysis, compliance validation against regulatory schemas, and AI-generated narrative explanations for quarter-over-quarter variations in financial ratios are capabilities that will change how statutory reporting teams operate. The patent's inclusion of a materiality indicator and historical query repository suggests a system designed to augment, not replace, actuarial judgment, but the line between augmentation and automation tends to shift over time.

Second, the Insurance LLM patent's explicit naming of actuaries as target users for guidance artifacts suggests that EXL sees claims-adjacent actuarial analysis (reserve adequacy assessment, claim leakage detection, severity trending) as a near-term deployment target for domain-specific AI.

Third, the breadth of EXL's patent portfolio (spanning document extraction through knowledge graphs through domain-specific LLMs through regulatory reporting) suggests a strategy of controlling the full analytical pipeline from raw data to actionable insight. For carriers that outsource significant operational functions to EXL, this concentration of patented capabilities in a single provider creates dependency risk that boards and CROs should evaluate.

Fourth, the three-way patent dynamic between carrier IP (AIG), vendor IP (Quantiphi), and services company IP (EXL) means that the "build vs. buy vs. outsource" decision for AI capabilities now includes an IP dimension. Carriers building internal systems need to navigate around all three patent portfolios. Carriers licensing vendor platforms need to understand how vendor IP interacts with services company IP. And carriers outsourcing to services companies like EXL need to understand what IP dependencies they are creating.

Patterns we have seen across the AIG, Quantiphi, and now EXL patent portfolios suggest that the AI patent landscape in insurance is still in its early stages. These are foundational patents covering first-generation architectures. As the technology matures, expect to see continuation patents that narrow and extend these claims, international filings that expand geographic coverage, and inevitably, freedom-to-operate challenges as competing architectures bump against overlapping claims.