Insurer AI patent filings looked, in aggregate, like a slowdown. The 166 patents filed by 30 major North American and European insurers between January 2023 and October 2025 put cumulative volume roughly 30% below the 2020 peak that preceded the generative AI era. But inside that aggregate, the composition shifted sharply. Generative AI patents grew from 4% of all insurer AI filings in 2014 to 31% as of October 2025, according to Evident’s Insurance AI Patent Tracker, published in December 2025. State Farm (326 total filings since 2014), USAA (218), and Allstate (136) held 77% of the entire tracked portfolio. P&C carriers accounted for 89% of all insurer AI patents. The trade press treated the Evident report as a carrier ranking exercise. The patent composition data is doing something more useful: it maps the boundary between what insurers are willing to disclose publicly and what they are not.

That boundary does not run where most observers expect. Carriers have spent the past three years telling analysts, regulators, and investors that underwriting and pricing represent their primary GenAI investment destinations. The Evident data shows that GenAI patents cluster in customer service and claims automation, not underwriting or pricing. This divergence between stated investment priority and protected IP reflects a deliberate calculation about which AI innovations benefit from public patent disclosure versus which ones require trade secret protection. The implications for competitive differentiation across carrier tiers run deeper than any headline filing count.

Why P&C holds 89% of all insurer AI filings

The 89% P&C share is not simply a function of market size. P&C carriers hold a structural patent filing advantage rooted in the nature of their AI applications. Telematics systems, sensor-based risk monitoring platforms, usage-based insurance architectures, aerial and satellite imagery processing for property damage, and IoT-driven underwriting triggers all share a common characteristic: they describe concrete physical sensing and data collection methods that satisfy the technical improvement requirement under 35 U.S.C. Section 101. A patent claiming a method for combining accelerometer, gyroscope, and GPS data to generate a real-time driving behavior score describes a specific technical system with a measurable hardware integration. Patent examiners can apply the Alice two-step test and find a genuine improvement in sensor data processing. That claim survives.

Life and health insurer AI applications are harder to frame this way. Mortality prediction models, morbidity trend algorithms, medical cost projection systems, and prescription drug utilization forecasters all operate primarily on statistical inference against administrative data rather than on technical hardware control. The claims risk characterization as abstract mathematical methods applied to insurance data, which is exactly the kind of application the Federal Circuit invalidated in Recentive Analytics, Inc. v. Fox Corp. in April 2025. When the Supreme Court denied certiorari in December 2025, that precedent became binding and final. Life and health AI developers face a genuine Section 101 problem that P&C developers have navigated more easily by grounding patent claims in the physical sensor architecture rather than the statistical model on top of it.

This explains why State Farm’s 326-patent portfolio concentrates so heavily in P&C and why USAA’s 218 filings, despite USAA’s multiline product range, anchor in property and auto applications. The patent-eligible surface area for P&C AI is structurally broader, which creates a compounding advantage: more filings generate more prior art, which then defines the design space that latecomers must navigate around.

The GenAI composition shift: where 31% actually lands

The growth from 4% to 31% GenAI patents in a decade reflects real technology adoption, but the composition of that 31% carries more information than the percentage. Evident’s tracker data shows GenAI filings concentrated in two areas: customer service applications (chatbots, virtual assistants, inbound query classification, policy explanation tools) and claims processing (FNOL summarization, damage description parsing, settlement communication generation). The overall patent portfolio has traditionally skewed toward claims and underwriting, where traditional ML built out most of its IP base, with the insurance sector filing over 300 AI patents in those categories, more than twice the volume of customer service, risk modelling, and pricing combined. The GenAI layer did not simply extend that map into newer model architectures. It opened a new front in customer interaction, the domain where generating text is the primary output and where Section 101 eligibility is easiest to establish.

Consider the contrast between two GenAI patent framings. A patent on a system that produces personalized policy renewal communications in natural language, drawing on claims history, coverage terms, and loss event context, describes a concrete output-generation method tied to a discrete technical function: language model inference conditioned on structured insurance data. That framing has a reasonable shot under current eligibility standards. A patent on a system that generates better initial loss reserve estimates by synthesizing unstructured claims notes with structured exposure data is, in effect, claiming a better statistical actuarial judgment. The second framing faces much higher Section 101 rejection risk and carries an additional cost that the first does not: revealing your reserving methodology to every competitor who reads the published application.

This is why the GenAI patent composition skews toward customer interaction rather than core actuarial functions, even though carriers invest heavily in both. The underwriting and reserving AI investment is real. It does not appear in the patent record because the carriers making those investments have decided, rationally, that disclosure is not worth the legal protection for systems that operate entirely inside their own infrastructure.

The 2020 peak and what the post-peak decline signals

Insurer AI patent filings peaked in 2020 and have not recovered despite the generative AI investment surge of the past three years. Current volumes sit approximately 30% below that peak. The standard narrative frames this as a prosecution lag: AI adoption accelerated, patent examination takes 18 to 36 months, and the filing wave for 2023 and 2024 GenAI investments has not yet cleared examination. That lag explanation is partially correct but not complete.

Three additional factors explain more of the gap. First, the vendor platform shift. A growing share of carrier AI capability since 2023 was deployed on licensed infrastructure: Guidewire ProNavigator, Palantir Foundry, Duck Creek’s Agentic AI Platform, Microsoft Copilot Studio. When a carrier deploys AI on a vendor platform, the patentable architecture typically belongs to the vendor. The carrier contributes training data, domain configuration, and workflow expertise but does not own the underlying model orchestration technology. You cannot patent a workflow you built on Palantir’s infrastructure without, in effect, describing Palantir’s platform. Second, the Section 101 quality filter. Post-Recentive Analytics, patent counsel at major carriers have grown more selective about which AI methods are worth the prosecution investment. Filing a patent that an examiner rejects under Section 101 wastes prosecution budget and signals to competitors that you attempted to protect a particular method without succeeding. Selective filing produces fewer but stronger claims. Third, trade secret preference. For AI systems operating entirely within carrier infrastructure, where no external party can observe or reverse-engineer the method, trade secret protection is often the rational alternative. The Defend Trade Secrets Act of 2016 strengthened federal enforcement. Carriers acting on that calculus file fewer patents, not because their AI programs are weaker, but because they have determined that disclosure does not justify the protection.

State Farm, USAA, and Allstate: what 680 patents buy

The Big Three’s combined 680 patents across the tracked period, approximately 77% of a total portfolio implying roughly 883 total insurer AI filings since 2014, deliver three distinct competitive assets beyond the patents themselves. Freedom to operate: a carrier with an extensive patent portfolio can deploy AI systems without meaningful risk that a competitor or patent assertion entity will bring an infringement claim in a technology domain it has already documented. Licensing leverage: a portfolio of 326 patents gives State Farm substantial negotiating power in any technology partnership, joint venture, or M&A process where AI IP is being valued. M&A signal quality: a portfolio this large and consistent provides acquirees and acquirers alike with a documented, examinable record of technical development that trade secrets cannot replicate during due diligence.

Carrier AI Patent Filings (Since 2014) Approximate Share
State Farm 326 ~37%
USAA 218 ~25%
Allstate 136 ~15%
All other tracked carriers (27) ~203 ~23%
Total tracked ~883 100%

Each of the Big Three has built its portfolio through sustained filing programs reflecting genuine technical development capacity. State Farm’s filings span telematics processing, claims triage automation, agent-facing digital tools, and more recently GenAI-assisted customer communication. USAA’s portfolio concentrates in property damage assessment through its aerial imagery generative AI work, auto claims processing for members deployed overseas, and the multi-agent coordination architectures that make USAA the only insurer with meaningful agentic AI patent depth. Allstate’s filings include AI drift detection and model monitoring patents covering the governance control layer, a sophisticated framing that sidesteps Section 101 concerns by focusing on system oversight rather than prediction quality.

The remaining 23% is distributed across approximately 27 other carriers in Evident’s tracked universe. The per-carrier average outside the Big Three works out to fewer than eight patents across a 12-year period. That is not a patent program. It is occasional opportunistic filing around discrete innovations, without the sustained prosecution infrastructure that produces competitive IP density.

Customer service over underwriting: the strategic intent gap in the filing record

Carrier earnings presentations and investor day materials from 2023 through early 2026 have been consistent in framing GenAI investment priorities: underwriting efficiency, pricing precision, and risk selection quality as the primary destinations. State Farm’s OpenAI Frontier deployment emphasizes Navi for agents but carrier leadership has been clear that long-term value lies in underwriting and claims. Travelers deployed Anthropic to 10,000 engineers across actuarial and engineering workflows. AIG’s Palantir partnership explicitly targets underwriting automation across a $1.6 billion specialty portfolio. These are real programs with real capital behind them.

The GenAI patent portfolio does not reflect this emphasis. The patents that have cleared examination and entered the public record concentrate in customer service: policy inquiry response generation, renewal communication personalization, FNOL guided interview automation, and claim status update production. The underwriting and pricing applications of GenAI, described by carrier leadership as primary investment areas, are largely absent from the patent record.

The explanation is not a mismatch between stated and actual investment. The explanation is that underwriting and pricing AI is being protected through trade secrets, while customer service AI is being patented. This allocation is rational. A customer service AI system, once deployed in customer-facing channels, can potentially be observed and reverse-engineered by competitors monitoring interaction patterns. Patent protection provides exclusive rights against independent parallel development. An underwriting pricing model, by contrast, operates entirely inside carrier systems. No competitor can directly observe how it weights variables, selects training features, or generates price indications. Trade secret protection is adequate and carries no disclosure cost. The result is a patent record that systematically understates how advanced carrier underwriting AI actually is, because the most strategically sensitive work is protected in ways that leave no public record.

This creates an information asymmetry with regulatory implications. NAIC’s multistate AI evaluation pilot, running through September 2026 across twelve participating states, is designed to give regulators a structured framework for examining AI systems during market conduct reviews. A carrier whose customer service AI is patented has already produced a detailed technical specification in the patent application itself. That documentation maps directly onto what examiners need: a description of inputs, outputs, decision pathways, and system architecture. A carrier whose underwriting AI is protected as a trade secret faces a harder compliance disclosure problem: providing the architectural detail regulators want risks degrading the trade secret protection if that documentation enters a public regulatory record.

The agentic frontier: three filers and deepening concentration

The concentration pattern visible in the GenAI layer is already repeating at the next technological frontier, and the numbers are more extreme. Only three insurers have filed agentic AI patents at all. USAA leads with multi-agent coordination claims focused on aerial imagery-based property damage assessment and adaptive underwriting decision chains with continuous feedback loops. The concentration at the agentic layer is tighter than even the GenAI layer in 2014, when at least a handful of carriers were filing GenAI-adjacent claims. The agentic category is forming its prior art base with essentially one dominant filer.

Evident’s Q4 2025 use case report, published separately from the patent tracker, found that agentic AI systems accounted for 21% of publicized insurer AI deployments in Q4 2025, with claims management representing 37% of those agentic projects and five insurers actively testing agentic AI in claims. The gap between deployment volume and patent activity is wider at the agentic layer than anywhere else in the tracker data. Carriers are running agentic systems in production while filing almost no agentic patents. Evident anticipates this will change in 2026, with expected activity concentrated in system-level designs, multi-agent coordination, control mechanisms, and continuous feedback loops.

The first patents filed in any AI subcategory define what counts as obvious versus novel in subsequent examinations. USAA’s early agentic filings are establishing design choices, coordination protocols, and governance architectures as prior art before most carriers have drafted their first agentic claims. The full analysis of USAA’s agentic patent portfolio covers the specific multi-agent coordination and feedback-loop claims in detail. What the Evident aggregate data reveals is that this concentration dynamic is structural, not accidental: it repeats at each new frontier because the same carriers have the patent prosecution infrastructure to move first.

Mid-market carriers and the vendor IP calculus

For the roughly 27 carriers outside the Big Three in Evident’s tracked universe, the patent gap presents a strategic question that extends beyond IP counts. The question is whether mid-market carriers are conceding the patent dimension of the AI race in favor of vendor-led deployment, and what that means competitively if vendor platforms commoditize the AI functionality that current filings are designed to protect.

The commodity trajectory is plausible. Guidewire ProNavigator, Duck Creek’s Agentic AI Platform, and Verisk’s MCP connectors launched within six weeks of each other in early 2026, each targeting the embedded AI layer in P&C core systems. If a mid-market carrier can deploy production-grade underwriting AI by subscribing to Guidewire’s platform, the case for investing in a proprietary patent program weakens considerably. The carrier’s AI capability is real and operational, but the architecture IP belongs to Guidewire. The carrier builds workflow expertise and training data advantages rather than patent moats.

The risk in this strategy surfaces during consolidation. When a mid-market carrier becomes an acquisition target, technology due diligence will ask what AI capability the carrier owns versus rents. A carrier with eight patents and extensive Guidewire configuration expertise cannot demonstrate owned IP comparable to State Farm’s 326-patent portfolio. The operational AI is real, but the premium that acquirers pay for proprietary technology is unavailable. The acquisition premium concentrates in book of business, distribution relationships, and geographic footprint rather than technology assets. For carriers that view AI investment as a component of enterprise value rather than just operating efficiency, the build-vs.-buy decision has a patent valuation dimension that vendor-led strategies structurally cannot optimize. The AM Best survey quantifying that 68% of carriers outsource AI while only 18% actively track vendor risk describes precisely the governance gap that this ownership question eventually produces.

The 2025 Section 101 reset and what it opens for 2026 agentic filings

The Trump administration’s January 2025 Executive Order on AI, titled “Removing Barriers to American Leadership in Artificial Intelligence,” directed federal agencies to review and revise any Biden-era AI policies that constrained American AI leadership. At the USPTO, this produced a November 2025 guidance revision that rescinded substantial portions of the February 2024 Biden administration AI guidance, which had imposed stricter inventorship standards for AI-assisted inventions and created additional eligibility scrutiny for AI-method claims. The USPTO simultaneously advanced Ex Parte Desjardins as a precedential PTAB decision, providing a clearer eligibility pathway for claims demonstrating concrete technical improvements to the functioning of an AI system rather than merely applying known AI methods to a business context.

Combined, these shifts make the 2026 examination environment meaningfully more permissive than what carriers encountered in 2024 and early 2025. For agentic insurance patents specifically, claims that describe technical improvements to multi-agent coordination protocols, adaptive feedback loop architectures, or control mechanisms governing autonomous decision thresholds now have a cleaner path to grant than they did 18 months ago. The Section 101 analysis that previously threatened to invalidate agentic claims as abstract ideas applied to insurance processes has been reoriented toward asking whether the claimed system improves the functioning of the AI system itself. For well-drafted agentic claims, the answer is yes.

For carriers with agentic systems in production, the current window carries real strategic weight. Filing now, while the category is nearly empty and the examination environment is favorable, builds a prior art foundation that constrains every subsequent filer. The carriers that did not file telematics patents in 2013 and 2014 are still navigating around State Farm’s prior art positions a decade later. The same dynamic is forming at the agentic layer, and the window for establishing foundational positions is short. Our detailed analysis of the Section 101 reset covers the examination implications and the specific claim structures that the Desjardins framework supports.

The post-Biden guidance rollback also affects GenAI applications already in prosecution. Patent applications filed in 2023 and 2024 that were pending under the stricter Biden-era framework may now clear examination under the revised guidance, adding to the published portfolio and expanding the scope of documented insurer-owned AI IP. Some of the filing volume suppression of the past two years may resolve into grants over the next 12 to 18 months as backlogged applications work through the system.

What the patent map signals for actuarial practice

The Evident patent data, read carefully, gives actuaries a map of where autonomous decision-making in insurance has public IP backing and where it does not. That distinction carries practical consequences for model validation, rate filing documentation, and risk assessment work.

Customer service and claims AI, the two categories with the deepest GenAI patent coverage, are also the domains where NAIC’s AI governance framework is most developed. A carrier whose claims AI operates under a patented architecture has already produced, in the patent application itself, a technical specification of how the system works, what inputs it processes, what outputs it generates, and what decision boundaries govern autonomous versus human resolution. That documentation maps onto what regulators need during an AI system examination under the NAIC Model Bulletin on AI, enacted in approximately 24 states. The patent does not constitute regulatory approval. It does provide a technical baseline that trade-secret-protected systems cannot deliver without potentially compromising the secrecy that trade secret law requires.

For underwriting and pricing AI, which the patent record mostly omits, actuaries operating under ASOP No. 56 face a different challenge. The models are real and consequential, generating pricing, reserving, and risk selection outputs that flow through to policyholder outcomes. But the architecture is not publicly documented. When an appointed actuary opines on the reasonableness of reserves generated by a system that uses proprietary GenAI-assisted reserving tools, the validation documentation must stand on internal methodology papers rather than publicly examinable technical specifications. The opacity that trade secret protection provides in the competitive market also limits the information available for external validation.

The agentic patent gap compounds this. Only three insurers can point to a public record of their agentic AI architecture. The rest are deploying agentic systems in production without any public documentation of how those systems coordinate decisions, what control mechanisms govern autonomous actions, or how feedback loops adapt agent behavior over time. For actuaries asked to validate the reasonableness of reserving and pricing outputs from systems they cannot directly audit, the gap between what carriers deploy and what appears in the patent record shapes the information available for that work. The carriers that chose to patent their agentic architectures will find actuarial validation and regulatory examination more tractable. The documentation that patent prosecution requires is precisely the documentation that ASOP No. 56 calls for in model governance frameworks, and the gap between the two is not going to close on its own as these systems scale.

Further Reading

Sources