Patra's AI Patent Signals a New Tier of Insurance IP Players

Patra Corporation, a privately held insurance outsourcing firm, was granted U.S. Patent No. 12,639,972 on May 26, 2026 for an AI policy-checking method that requires a mandatory human confirmation step before extracted values are accepted, a claim structure built to survive USPTO's tightened Section 101 guidance and a sign that insurance AI patent activity is spreading beyond the five carriers holding more than 900 patents (CB Insights, 2026).

What Patent 12,639,972 Actually Claims

The patent, titled "Systems and Methods for Value Extraction and Guided Review," names Patra Chief Technology Officer Tony Li and Senior Director of AI Engineering Juan Cristian Martinez Vega as inventors, with Patra Corporation as sole assignee (USPTO Patent No. 12,639,972, May 26, 2026). Its priority date traces back to February 2022 (Patra, July 2026), meaning the application sat in prosecution for roughly four years while the USPTO's own standards for AI subject-matter eligibility shifted twice underneath it. The claims describe a system that scans an insurance document, maps the position of words on the page, locates a target data field such as a coverage limit or an effective date, identifies the value associated with that field, and then routes the result to a human reviewer for confirmation before it is accepted into the record. Accuracy improves over time through a feedback loop centered on that expert review step, not through unsupervised model retraining alone.

That workflow underlies Patra's commercial Policy Checking AI product, which the company markets to retail agencies, wholesalers, MGAs and MGUs, and carriers as a way to automate the traditionally manual task of verifying that an issued policy matches what was quoted and bound. Patra's chief executive, Pratap Sarker, framed the grant as validation of a design choice the company made years before regulators caught up to it: "We were building governed, expert-supervised AI with real review at the center of the process before that became the industry's accepted position" (Patra, July 2026). Vega, one of the named inventors, put it more narrowly: "This patent illustrates the foundation of the AI and human-in-the-loop mechanisms that are still at the core of what Policy Checking AI is today" (Patra, July 2026). Patra says the product is backed by market-leading errors-and-omissions coverage, an unusual detail for a workflow-automation announcement and one that points directly at the liability question underneath the patent, addressed further below.

The Human-Confirmation Step as a Section 101 Survival Strategy

Patent claims that amount to "apply artificial intelligence to an existing business process" have had a rough run at the USPTO and in the courts since the Federal Circuit's 2025 Recentive Analytics decision narrowed what counts as an eligible AI invention. On August 4, 2025, the USPTO issued a memo that rescinded the more permissive February 2024 guidance and pulled examiners back toward the traditional Alice/Mayo framework, cautioning them against stretching the "mental process" exclusion to cover computational steps that genuinely cannot be performed in a human mind, but also instructing that a Section 101 rejection should issue only when it is more likely than not that a claim is ineligible (USPTO, August 2025). That is a narrower, more examiner-discretion-dependent standard than insurers filed against for most of the 2020s, and it has already reshaped how carriers and vendors draft claims, a shift documented in this site's coverage of the broader Section 101 reset and in the Federal Circuit rulings that followed it.

Patra's claims read like a direct response to that environment. A claim that simply says "extract a value from a document using AI" is the kind of abstract-idea-plus-generic-computer claim examiners are now trained to reject. A claim that specifies scanning the document, mapping word positions on the page, locating a target field, extracting an associated value, and then routing that value through a defined interface for a human reviewer to confirm or override before it is committed describes a concrete technical pipeline with a non-optional human determination as its final step. That structure does two things at once: it ties the abstract idea of "checking a policy" to specific technical means, and it places a human judgment call, not an algorithmic output, at the point where the claim's practical result is finalized. Whether Patra's drafting team wrote the claims this way in 2022 anticipating the eligibility fight to come, or amended them during prosecution as the standard shifted, the outcome is a patent that survived examination in an environment openly hostile to thinner AI-automation claims.

Where a Vendor Patent Sits Against the Five-Carrier Concentration

Set against the scale of carrier patent portfolios, one BPO vendor's single grant looks modest. CB Insights' AI Readiness Index found that just five companies, State Farm, UnitedHealth Group, Allstate, Cigna, and The Hartford, account for more than 900 of the US AI patents filed by the insurers it evaluated, with State Farm alone holding over 300 patents including more than 50 in computer vision (CB Insights, 2026). A separate Evident tracker, covering 30 major carriers, found State Farm, USAA, and Allstate specifically hold 326, 218, and 136 AI patents respectively, a combined 77% of all insurer AI filings since 2014 (Evident, December 2025), a concentration this site examined in detail in its analysis of the three-carrier patent gap.

The volume gap is real, but it obscures where each tier is actually competing. Carriers patent broadly across functions they run in-house: claims triage, pricing, fault determination, and underwriting decisioning. None of the five carriers in the CB Insights count are in the business of manually rekeying endorsement data for retail agencies and wholesale brokers, because that work sits downstream of the carrier, inside the distribution and outsourcing layer served by firms like Patra, Zywave, and Ebix. A single, well-drafted patent covering a specific administrative workflow in that layer does not need hundreds of siblings to matter commercially. It only needs to cover a task performed at scale across an 880-plus-client base (Patra, 2026), and to survive the same eligibility scrutiny that is now killing thinner claims from much larger filers.

Why Policy-Checking Accuracy Is an Upstream Data-Quality Problem

Policy checking is the quality-control step that catches transcription errors between what a broker quoted, what an underwriter bound, and what actually appears on the issued policy: a wrong coverage limit, a misstated effective date, an endorsement schedule that dropped a location. When that check is manual, error rates scale with reviewer fatigue and submission volume. When it is automated but unsupervised, an AI model's own extraction error can slip through with no human ever looking at the document again, which is precisely the failure mode Patra's claim structure is designed to prevent by forcing a confirmation step before any value is accepted.

Actuaries rarely model policy-checking accuracy directly, but the errors it catches or misses do not stay contained to the file where they occur. An uncaught limit or effective-date error does not just create a contractual dispute if a claim is filed against the wrong terms; it also propagates into the bound-policy data that eventually feeds loss development triangles, exposure bases, and rate filings for the MGA or wholesaler's book of business. A transcription error that understates an insured's limit, for instance, will make a subsequent large loss on that policy look like an outsized severity outlier relative to the (incorrect) exposure recorded, quietly distorting frequency and severity trends for that segment months or years before anyone traces the discrepancy back to a policy-checking gap. For MGAs and wholesalers specifically, the liability exposure is more immediate: errors and omissions claims against program administrators frequently originate in exactly this kind of documentation mismatch between what was quoted, bound, and issued, which is why Patra pairs its automated tool with its own E&O coverage rather than leaving the liability question to the carriers and retail agents downstream.

Freedom-to-Operate Risk for Carriers Building This In-House

A granted patent on a specific guided-review workflow changes the calculus for any carrier or MGA whose internal automation team has built, or is planning to build, a similar tool rather than licensing one. Insurance software has generally treated document-extraction-plus-human-review as generic, off-the-shelf functionality, the kind of thing an internal data science team assembles from open-source OCR libraries and a review queue without much IP diligence. A patent that claims the specific combination of document scanning, word-position mapping, target-value extraction, and mandatory human confirmation as a unified method narrows that assumption considerably, at least for anyone replicating the same sequence of steps rather than designing around it.

The pattern of specialist patent holders asserting rights over what carriers assumed was generic infrastructure is not hypothetical. The Hartford filed a declaratory-judgment suit against Intellectual Ventures in the US District Court for the District of Delaware on April 7, 2026, seeking a ruling that it does not infringe eight Intellectual Ventures patents and that those patents are invalid. The patents at issue target The Hartford's use of Docker, Kubernetes, Apache Spark, Apache Airflow, MongoDB, and Elasticsearch, open-source and third-party tools used broadly across the technology industry, with the infringement theory resting entirely on off-the-shelf products rather than anything Hartford built or customized itself (Insurance Business, April 2026). The filing states Intellectual Ventures has pursued similar claims against at least three other major insurers, Nationwide, Liberty Mutual, and GEICO, a pattern this site covered in its analysis of carriers' preemptive suits against patent-licensing firms. Patra's patent is a different animal, a practicing entity asserting its own product IP rather than a licensing firm targeting embedded infrastructure, but it establishes the same underlying exposure: once a specific technical workflow is patented, any carrier or vendor replicating that exact pattern internally, without a license or a design-around, inherits infringement risk it did not have when the workflow was assumed to be generic.

Human-in-the-Loop Is Becoming the Default Insurance AI Patent Claim

Patra's claim structure is not an isolated adaptation. Duck Creek's Agentic AI Platform, launched in April 2026 and expanded through its July 2026 acquisition of underwriting orchestration vendor Send, is built around what the company's own executives describe as augmentation rather than replacement of the underwriter, a framing this site examined in its coverage of the Send acquisition. EXL's $310 million acquisition of iMerit in June 2026 bought a human annotation and expert-review layer specifically to strengthen the training and evaluation side of its AI stack, detailed in this site's analysis of that deal. Across product launches, acquisitions, and now patent claims, insurance AI vendors are converging on the same design and drafting pattern: keep a human determination inside the loop, both because regulators are asking for it and because the USPTO's post-2025 eligibility standard rewards claims that tie an abstract automation idea to a concrete, non-optional human step.

That convergence sets up the next likely front in insurance AI patent litigation. Intellectual Ventures has already shown that a specialist patent holder will target the plumbing underneath a carrier's technology stack rather than the carrier's own proprietary models. As more vendors patent specific orchestration and guided-review workflows the way Patra just did, carriers and larger BPO competitors building comparable in-house tooling become plausible targets for the same kind of assertion, not from open-source infrastructure patents this time, but from operating-company patents covering the exact human-AI handoff sequence a growing share of the industry is standardizing around.

Why This Matters

For pricing and reserving actuaries, the lesson is not that Patra's single patent reshapes the competitive landscape. It is that the quality controls a BPO vendor builds into its policy-checking workflow are a direct input into the reliability of the bound-policy data an MGA's book eventually generates, and a patent grant is a signal, however indirect, about how seriously a vendor has engineered that control. Actuaries evaluating an MGA or wholesaler's data quality should ask what automated checks sit between quote, bind, and issuance, and whether a human reviewer genuinely confirms extracted values or merely rubber-stamps an AI output under production-volume pressure. For carriers and larger outsourcing firms building similar tooling internally, the practical takeaway is to run a freedom-to-operate check against workflow-specific patents like Patra's before scaling an in-house guided-review system, rather than assuming document-extraction-plus-human-confirmation remains unpatented generic infrastructure. The innovation frontier in insurance AI patents is no longer just the five carriers with the biggest portfolios. It now includes the specialist vendors underneath them, and the claim language they are writing to survive an examiner corps that has gotten considerably less forgiving.