Insurance AI patent claims are pivoting from single-function classifiers to system-level, multi-agent architecture, and the shift raises the drafting bar in a way that likely deepens rather than dissolves the existing patent concentration. State Farm, USAA, and Allstate already hold 77% of the 300-plus AI patents insurers filed on claims and underwriting (Evident, December 2025), and system-architecture claims demand a documentation depth that only a handful of carriers currently have the legal infrastructure to produce.

Patterns we’ve tracked across USPTO filings from the top property and casualty carriers over the past 18 months show a clear inflection point: claim language is moving from “a model that detects X” to “a system that coordinates N agents to resolve Y,” and that shift changes who can defensibly patent what. Evident’s CEO, Alexandra Mousavizadeh, put the stakes plainly around the firm’s tracker data: “Either patents remain the domain of a few frontrunners, or they become merely a signal of broader competitive intent” (Evident AI Patent Tracker, December 2025). The claim-language pivot now underway is what will decide which of those two futures the industry gets.

The bridge patent: USAA’s classifier claim meets the orchestration era

USAA’s U.S. Patent 11,810,347, “System and Method for Assessing Damage from Aerial Imagery,” filed November 5, 2021 and granted November 7, 2023, is the clearest bridge case between the two eras of insurance AI patenting. The claims describe a linear pipeline: an aerial system captures images of a known structure, a ground system crops those images around the selected location, and a damage classifier automatically categorizes the severity of structural damage. That is a single-function claim in the classic sense. Input goes in one end (raw aerial imagery), a model processes it, output comes out the other end (a damage classification), and the patent protects that specific input-to-output chain.

Claims of this shape have been the dominant style across the industry’s 300-plus AI patents on claims and underwriting, twice the filing volume of the next most common focus areas, customer service and risk modeling and pricing combined (Evident, December 2025). They are also, by comparison to what comes next, relatively straightforward to draft and to defend. A classifier claim maps cleanly onto the kind of technical improvement patent examiners look for: a specific model architecture applied to a specific data type to produce a specific, verifiable output. USAA holds 218 AI patents overall (Evident, December 2025), and the aerial-damage classifier sits comfortably within that portfolio’s dominant pattern.

What is entering the pipeline now is architecturally different. Evident expects 2026 filing activity to pivot toward “system-level designs, multi-agent coordination, control and continuous feedback loops” as insurers formalize production agentic use cases into patentable form (Evident AI Patent Tracker via Insurance Journal, December 2025). A claim covering a single damage classifier and a claim covering an orchestration layer that routes work across five or seven specialist agents are not variations on a theme; they require an entirely different approach to claim drafting, and they face a materially different eligibility analysis at the USPTO.

What a multi-agent orchestration claim must cover to survive Section 101

The gap between the two claim styles is not just descriptive length. Under the Alice/Mayo framework that examiners and courts apply to Section 101 eligibility, a claim first gets tested for whether it is directed to an abstract idea, and if so, whether it adds an “inventive concept” sufficient to transform that abstract idea into something patent-eligible. A single-function classifier claim like USAA’s aerial-damage patent survives this test relatively easily: it ties a specific technical process (capture, crop, classify) to a physical output (a damage assessment) using a defined data pipeline. The Federal Circuit has been unambiguous about what does not survive it. In Recentive Analytics, Inc. v. Fox Corp. (April 2025), the court held that merely applying known machine learning methods within a new data environment does not clear the Section 101 bar, and the Supreme Court’s denial of certiorari in December 2025 cemented that as binding precedent.

Multi-agent orchestration claims face the same test but a harder version of it. A claim that simply recites “a plurality of AI agents coordinating to process an insurance claim” is almost certain to be characterized as the abstract idea of delegating a business process to software, dressed in agent terminology. To survive, the claim has to identify a concrete technical improvement in how the coordination itself works, not in the business outcome the coordination produces. That means specifying the actual mechanism: what data structure agents use to hand off state to one another, what triggers an escalation from autonomous resolution to human review, and how a feedback signal from a completed case measurably changes the confidence threshold governing the next case. The Federal Circuit’s April 2026 decision in In re Brian McFadden reinforced how narrow the path is, affirming a rejection where the application enhanced abstract calculations on standard hardware without defining how the implementation physically or logically altered computer operation. A multi-agent claim that cannot point to an analogous, concrete alteration of how the system itself operates, not just what business result it achieves, will meet the same fate.

That is a substantially higher drafting burden than a classifier claim. It requires the patent applicant to document, with specificity a court will accept, the internal control logic of a system that many carriers currently treat as an operational black box optimized for uptime, not for patent prosecution. The USPTO’s November 28, 2025 guidance reset narrowed the aperture further in one respect and widened it in another: it rescinded the Biden-era framework that let AI systems factor into inventorship analysis, directing examiners to apply Alice/Mayo uniformly and anchor the inventive concept in human effort, with no special pathway for AI-assisted inventions. For carriers, the practical upshot is that the technical narrative in the claim, not the sophistication of the underlying model, is what will determine whether a system-architecture patent issues at all.

Does the higher drafting bar widen the moat or open a new door

The concentration numbers raise an obvious question once the claim-language shift is on the table: does a harder-to-draft claim style deepen the existing 77% concentration among State Farm, USAA, and Allstate, or does it create room for a carrier outside that trio to leapfrog into a category that is, for now, nearly empty? Evident’s tracker found that only three insurers have filed agentic AI patents at all, with USAA leading that subcategory specifically. Two forces point in opposite directions.

The case for widening: system-architecture claims require in-house or retained patent counsel who can translate control-loop engineering into Alice/Mayo-compliant claim language, plus an internal engineering culture that documents orchestration logic to litigation-grade specificity rather than leaving it in Slack threads and internal wikis. State Farm, USAA, and Allstate have already built that legal and documentation infrastructure across 326, 218, and 136 filings respectively. A mid-sized regional carrier attempting its first system-architecture filing is not just catching up on patent count; it is building the underlying claim-drafting capability from zero, at the exact moment the eligibility bar for that claim style is at its least forgiving point in a decade.

The case for narrowing: agentic system claims are new enough that no carrier, including USAA, has an entrenched prior-art position covering the specific coordination architectures now emerging in production, the way State Farm effectively defined the classifier-claim design space over a decade of filings. A carrier with strong technical documentation of a genuinely novel control mechanism, filed in the next 12 to 18 months before the category’s prior art thickens, could establish a claim position the incumbents do not yet hold. That window narrows every quarter; each new USAA filing in the agentic category raises the novelty bar for the next applicant by expanding what counts as already-known art.

On balance, the concentration case is the stronger one. The three incumbents are the carriers best positioned to move first specifically because system-architecture patenting is now as much a legal-documentation exercise as an engineering one, and legal documentation capacity compounds with filing volume in a way raw AI capability does not. The practical effect for everyone else is that the agentic patent window functions less like an open door and more like a fast-closing one.

Project Nemo as the working example: reserving and ULAE implications of orchestrated claims

Allianz’s Project Nemo, launched in Australia in July 2025, is the clearest public illustration of the architecture now entering the patent pipeline. A central planning agent coordinates six specialist agents, coverage verification, weather confirmation, fraud screening, payout calculation, and audit, to resolve eligible food-spoilage claims under $327 (Allianz, November 2025). The result is an 80% reduction in claim processing and settlement time, with in-scope claims resolved in under five minutes from filing to human final review, down from several days under the prior process.

That compression carries two concrete actuarial implications beyond the operational headline. The first is on unallocated loss adjustment expense. A carrier that automates the straightforward end of a claims segment does not simply cut ULAE uniformly across that segment; it strips out the low-complexity claims that previously absorbed a share of fixed adjusting overhead, leaving the claims still requiring human handling to carry a larger allocated share of that fixed cost per claim. Actuaries pricing or reserving for a line where an orchestrated system has gone live need to re-baseline the ULAE allocation between the automated and the still-manual tail rather than applying a blanket reduction factor, or the manual tail will look artificially cheap relative to its true burden.

The second is on loss development triangles. A claims segment where settlement time compresses from several days to under five minutes for a defined eligibility band is not a uniform acceleration of the existing payment pattern; it is a bifurcation of the segment into two populations with structurally different development. Mixing agentic-eligible and traditionally-processed claims into a single triangle risks distorting both the tail factor and the IBNR estimate, because the triangle will show an apparent acceleration in early-development payment percentages that reflects a change in claims mix, not a change in the underlying loss emergence pattern for claims still requiring adjuster judgment. The clean approach is to segment the triangle by processing pathway before the orchestrated system reaches enough volume to distort the blended pattern, which for a fast-scaling deployment can happen within two or three accident quarters.

Build versus buy: the calculus for mid-tier carriers

Not every carrier can fund a system-architecture patent portfolio, and the vendor landscape has organized around that reality. Starr Insurance selected Five Sigma’s multi-agent Clive platform for specialty claims operations, while James River Insurance partnered with Kalepa for AI-driven underwriting in the excess and surplus lines market, both examples of carriers accessing agentic architecture through a licensed platform rather than building and patenting their own. Global InsurTech investment reached $943.4 million across 42 deals in the first quarter of 2026, with roughly 75% of that capital flowing to AI-focused companies, and the average AI-focused InsurTech round hit $33.7 million against $14.2 million for non-AI rounds in the same quarter, evidence that capital is backing the vendor layer as the default distribution mechanism for agentic capability rather than betting broadly on carrier-built systems.

For a mid-tier carrier, the build-versus-buy math is not simply about upfront cost. A carrier that licenses an agentic claims or underwriting stack gets the operational benefit, the cycle-time compression and the ULAE relief, without the multi-year legal investment a defensible system-architecture patent requires. What it forfeits is IP ownership of the specific orchestration logic running its book. In an acquisition, that logic shows up in the diligence file as a vendor contract, not a patent asset, and it carries no exclusionary value against a competitor licensing the same platform. Carriers with genuinely unusual product logic, specialty-line complexity, or state-specific rating and compliance requirements also tend to hit vendor configuration boundaries faster than initial estimates project, at which point they are funding custom development at platform-licensing prices without the patent upside that would come from building the equivalent system in-house. The realistic mid-tier posture is a hybrid one: license the commodity orchestration layer for standard lines, and reserve any internal patent-filing budget for the narrow set of coordination mechanisms that are genuinely proprietary to how the carrier underwrites or adjusts its most distinctive business.

What a defensible actuarial pricing agent patent would need

Every public example of agentic architecture entering the patent pipeline, USAA’s claims and underwriting workflows, Project Nemo, Five Sigma’s Clive, sits on the claims or intake side. Underwriting and pricing lag noticeably in agentic patent volume, which is itself informative: a pricing agent that autonomously adjusts rate factors is a harder Section 101 case than a claims agent that autonomously resolves a $200 spoilage claim, because rate-setting is more directly entangled with an approved rate filing, a state regulatory record, and the actuarial standard of practice governing rate changes.

A defensible claim for an actuarial pricing agent would likely need to anchor its inventive concept in that regulatory entanglement rather than in the pricing calculation itself, since “an AI agent that computes an insurance rate” is squarely the kind of abstract business method the post-Recentive and post-McFadden eligibility landscape rejects. The more defensible architecture would describe a specific technical synchronization mechanism: a rating agent that proposes a factor adjustment only when a defined materiality threshold on incoming exposure data is breached, paired with a compliance agent that checks the proposed adjustment against the carrier’s filed rating plan in the relevant state before the change can commit, with the interaction between those two agents, not the rate math, as the claimed invention. That framing ties the patent to a concrete technical problem, keeping an autonomous pricing system synchronized with a filed regulatory record in real time, that has no direct analogue outside insurance, which is exactly the kind of domain-specific technical improvement that survived scrutiny in the USPTO’s November 2025 guidance where a generic optimization claim would not.

Why this matters

The claim-language shift from single-function classifiers to system architecture is not a drafting nuance; it resets who can plausibly hold defensible IP over the agentic workflows insurers are already running in production. Carriers deploying orchestrated systems without a parallel patent strategy are building operational advantages, faster cycle times, lower ULAE, that remain legally unprotected and open to replication by any competitor willing to license a comparable vendor stack. For actuaries, the more immediate consequence is methodological: wherever an orchestrated agent system goes live inside a reserved line of business, the loss development triangle and the ULAE allocation built for the pre-automation claims process no longer describe the business as it now operates, and both need to be re-segmented before the blended data quietly misstates the reserve.

Further Reading

Sources