Evident's Insurance AI Patent Tracker, released in December 2025, established that State Farm (326 patents), USAA (218) and Allstate (136) hold 77% of all insurer AI filings since 2014. Inside that aggregate sits a subcategory almost nobody occupies: only three insurers have filed agentic AI patents at all, and USAA leads it.

Generative filings went from 4% to 31% of the total between 2014 and October 2025. Agentic filings stayed rare while agentic deployment accelerated, and the gap says something specific about what carriers can actually own.

Key Takeaways

  • Three insurers have filed agentic AI patents while 22% of insurers plan agentic systems in production by the end of 2026, an adoption-to-IP ratio with no parallel in the generative category.
  • Insurance AI deployments grew 87% year over year in Evident's Q4 2025 data, with agentic systems accounting for 21% of publicized deployments and 56% of those aimed at claims management.
  • USAA's underwriting claims describe a five-agent chain ending in a decision orchestrator that routes to a human, then feeds the override rationale back to the upstream agents.
  • The Federal Circuit's April 2025 Recentive Analytics decision, with certiorari denied in December 2025, means a claim reading "apply multi-agent coordination to underwriting" fails Section 101 outright.
  • The NAIC Model Bulletin now sits in roughly 24 states, and the 12-state AI Evaluation Tool pilot runs from January through September 2026.

What an Agentic Claim Actually Covers

The distinction from a generative patent is architectural. Generative claims describe an input-output pipeline: a model receives data, processes it, returns a result. USAA's generative filing for clarifying aerial property damage imagery is the textbook case, enhancing and annotating storm imagery to speed catastrophe claims assessment.

Agentic claims cover system-level designs, and the filings that qualify carry three features together. Multiple specialized agents operate in concert, and the patentable innovation is how they pass structured outputs to one another rather than what any single model does. Explicit control mechanisms define what an agent may do autonomously and what forces escalation. Feedback loops run inside the workflow rather than in periodic offline retraining, so routing and escalation decisions adjust in near real time.

USAA's underwriting filings put that into a sequence. A submission intake agent normalizes application data across PDFs, spreadsheets and API feeds. A risk profiling agent cross-references it against guidelines and historical loss data. A pricing agent structures terms. A compliance agent checks rate filing and disclosure obligations. A decision orchestrator aggregates all of it and decides whether the case clears automatically.

The patentable feature sits at the end of that chain. When the orchestrator escalates and a human underwriter overrides the preliminary recommendation, the override rationale flows back to the relevant agents. The claims-side filings work the same way, tracking customer acceptance rates, reopened claims and litigation frequency, and adjusting the orchestration logic from them.

The Threshold That Moves Is an Actuarial Assumption

That feedback loop is the part with consequences beyond IP strategy, because what it recalibrates is the boundary between machine-closed and human-handled work.

Metric Value Source
Year-over-year AI deployment growth 87% Evident Q4 2025
Share of Q4 2025 deployments that were agentic 21% Evident Q4 2025
Agentic deployments focused on claims 56% Evident Q4 2025
Insurers planning agentic production systems by end of 2026 22% Industry surveys
Insurers with agentic AI patents 3 Evident Patent Tracker
Insurers reporting tangible AI business benefits 40% Evident Q4 2025

56% of agentic deployments target claims management, which is where the effect lands hardest. A conventional straight-through processing rule is static: claim types on the list close automatically, and the mix of what closes fast versus slow holds still between rule changes. An orchestrator that learns which claim types it can safely resolve, from reopened-claim and litigation-frequency feedback, moves that mix continuously and without a change control event.

Paid and incurred development factors are fitted on the assumption that the closure pattern by claim type is stable across the triangle. A system deliberately built to widen its own autonomy as evidence accumulates breaks that assumption from the inside, and it does so gradually, which is the hardest way for a diagonal to shift. The reserving question is not whether the automation is accurate. It is whether the settlement rate underlying the development pattern is being changed by a mechanism with no filing date attached to it.

The scale is already there. Hiscox compressed its London Market specialty quote cycle by 99.4%, from three days to roughly three minutes, escalating only final pricing. Allianz's Project Nemo cut food spoilage claims processing time 80% from first notice through settlement. AIG's Lexington subsidiary passed 370,000 submissions in 2025 against a 500,000 target for 2030.

The Category Is Empty Because Most of It Is Not Ownable

The obvious reading of three filings against 22% planned production is that carriers are slow. The likelier reading is that most of what they deploy cannot be patented by them.

A growing share of agentic deployment runs on vendor platforms. AIG's collaboration with McGill and Partners applies agentic AI across a specialty portfolio worth up to $1.6 billion in gross premiums written, built on Palantir's Foundry. The carrier supplies domain expertise, training data and workflow requirements; the coordination technology is the vendor's IP. That structure produces operational advantage and no patent.

What remains carrier-built then meets Section 101. Recentive Analytics established in April 2025 that applying known machine learning methods in a new data environment does not clear eligibility, and the Supreme Court denied certiorari in December 2025. A claim has to articulate a concrete improvement in how the agents coordinate or how the feedback operates, which requires technical documentation most insurers have never produced for their AI systems.

Trade secrets are the default fallback, and they carry a specific weakness here. They give no protection against independent invention, so a competitor building a functionally equivalent orchestration layer from public material has no liability. Trade secret litigation has grown, up 25% in filings within a year of the 2016 Defend Trade Secrets Act and over 1,200 cases in US courts during 2025 according to the Berkeley Technology Law Journal, but that protects against misappropriation, not duplication.

Regulation squeezes the same choice from the other side. The NAIC Model Bulletin, now enacted in roughly 24 states, requires documented governance proportionate to consumer risk, and the 12-state AI Evaluation Tool pilot running to September 2026 gives examiners a structured way to ask how a system reaches its decisions. Answering that in detail is exactly what erodes a trade secret. A patent has already made the architecture public and takes its protection from the grant instead, which is why the empty category may stay empty for reasons that have nothing to do with carrier appetite.

Further Reading

Sources