State Farm Patents a Pricing Model That Rewrites Itself

State Farm Mutual Automobile Insurance Company won U.S. Patent 12,694,432 B2 on July 28, 2026, for a system that runs a GPT model against its own rating predictions, writes the code fix for the pricing sub-model that missed, tests the rewrite in a simulated environment, and deploys the updated template. Claim 1 requires no human sign-off at any of those seven steps.

The grant lands seven weeks after actuary.info covered State Farm’s arrangement with OpenAI to retool pricing logic across 96 million policies, and it goes a step further than that partnership: it converts predicted-versus-actual monitoring, the check every ratemaking actuary already runs by hand, into patented, closed-loop software that can rewrite the rating engine without waiting for the next scheduled model build.

The filing sits inside a much larger position. State Farm holds more AI patents than any other insurer, 326 filed since 2014, more than USAA and Allstate combined, according to patent-tracking firm Evident, a concentration actuary.info analyzed in detail this spring. Most of that portfolio covers claims triage and autonomous-vehicle fault detection. US 12,694,432 B2 is different: it is the first of State Farm’s major AI grants aimed squarely at the machinery that sets the price rather than the machinery that pays the claim, and it lands just seven weeks after the carrier’s OpenAI Frontier partnership put pricing retooling on the record.

Inside Claim 1: A System That Diagnoses and Rewrites Its Own Rating Code

US 12,694,432 B2, titled “Large Language Modeling Systems and Methods for Building, Testing, and Validating a Predictive Model,” names three State Farm engineers, Michael W. Fields, Stephen Prevatt, and Jeff Stoiber, all based in Bloomington, Illinois, and assigns the patent to State Farm Mutual Automobile Insurance Company (US 12,694,432 B2, USPTO, July 28, 2026). Claim 1 describes a computer system with at least one processor programmed to run seven operations in sequence: detect an issue with a predictive sub-model that itself runs on a trained large language model to output a predicted value for a product based on historical data; execute a “GPT model” (the patent’s own generic term for a generative pretrained transformer, not a reference to a specific vendor’s product) trained to output the difference between that predicted value and the actual value recorded for the same product; feed the detected issue and the GPT model’s output into a second, distinct “code LLM” that writes a new model software template containing the code changes; generate a simulation environment built from a plurality of historical record data; execute the new template inside that simulation; and update the template based on the simulation’s output to produce one or more updated predictive sub-models.

Two language models do two different jobs in that sequence. One finds the pricing miss. The other rewrites the code that caused it. The claim language is explicit about the handoff: “input the one or more detected issues and the output of the GPT model into a code LLM configured to output a new model software template including one or more computer executable instruction changes to the one or more predictive sub-models based upon a comparison of the one or more detected issues and the output of the GPT model” (US 12,694,432 B2, Claim 1).

The patent’s drawings describe the hardware behind that language. A central server the filing calls the MTA computer device sits inside State Farm’s company network, wired to a locally trained LLM, an external trained LLM, internal and external data sources, the code LLM, a prompt engineering system, and a client device an underwriter or actuary can query. The detailed description states plainly what the system is for: it is “configured to operate on input data related to predictive pricing models including to build, simulate, and validate a predictive pricing model and/or sub-model for calculating insurance rates for an insurance product” (US 12,694,432 B2, USPTO, July 2026). That is not an underwriting triage tool or a claims chatbot. It is a rating-engine maintenance system, patented.

Dependent claims narrow what counts as a triggering “issue.” Claim 4 defines it as “a consistent variance or difference between a predicted value for replacing the one or more products outputted from an existing predictive sub-model and an actual value for replacing the one or more products” (Claim 4); claim 5 extends it to “a trend exceeding one or more thresholds, wherein the trend includes differences between different predictive sub-models and actual values” (Claim 5). Both describe a predicted-versus-actual backtest of the kind a pricing actuary already runs against a rating plan every renewal cycle, here wired directly into code that rewrites itself. Claim 8 has the GPT model output “a list of data elements identified as being related to the differences” when a user prompts it, and the specification names the candidate elements: claims history, vehicle history, prior insurance history, and public records. Claim 10 lets that GPT model be trained for a specific jurisdiction, and claim 9 lets the system generate a production model straight from the updated template. Claims 6 through 8 add optional features, a natural-language interface, an alert transmitted on detection, a prompt-driven review, but they are dependent claims. A system practicing only claim 1 satisfies the patent without any of them.

Claim 21 closes the loop further. It has the processor, “after generating the one or more updated predictive pricing sub-models, generate at least one predicted value using at least one of the one or more updated predictive pricing sub-models” (Claim 21), the first place in the document that names the target explicitly as pricing rather than the broader “products” language claim 1 uses. Read against claim 9, the patent describes an engine that can detect a pricing miss, rewrite the responsible code, validate the rewrite in simulation, push it to production, and start pricing new business against it, all inside the scope of a single system claim.

Two Provisional Filings and a Fourteen-Month Path to Grant

The idea has a longer paper trail than the July grant suggests. State Farm filed two provisional applications, 63/556,253 on February 21, 2024, and 63/641,773 on May 2, 2024, before consolidating them into non-provisional application 18/666,490 on May 16, 2024. The application published as US 2025/0265624 A1 on August 21, 2025, and issued fourteen months later as US 12,694,432 B2 on July 28, 2026, with 21 claims across six drawing sheets, examined by Ibrahim N. El-Bathy and prosecuted by Armstrong Teasdale LLP (US 2025/0265624 A1, USPTO, August 2025). The Patent Office classified the claims under both G06Q 30/0283, the code for price determination and estimation, and G06F 30/20, the code for design simulation and testing, a pairing that ties the pricing claim to a specific, technical validation mechanism rather than to a generic “apply AI to pricing” recitation. That is the kind of narrow drafting that actuary.info’s prior reporting on the post-Recentive Section 101 landscape found is surviving eligibility challenges that broader AI claims do not.

A Correction Loop That Used to Take a Development Cycle

Set the patent’s seven-step claim against how a rating model traditionally gets corrected, and the compression is the story.

Step	Traditional pricing-model correction	Patent’s automated loop
Detect the miss	Pricing actuary reviews a predicted-vs-actual report on a scheduled cycle	GPT model continuously flags a variance or trend exceeding a threshold (Claims 4-5)
Diagnose the cause	Actuary and analyst research drivers by hand	GPT model outputs a data-element list automatically on prompt (Claim 8)
Draft the fix	Actuary specifies the change; a developer codes it	Code LLM writes the new model software template (Claim 1)
Test the fix	QA and actuarial review in a staging environment, often weeks	System builds a simulation environment from historical record data and executes the template (Claim 1)
Approve the fix	Peer review, sign-off, rate filing where the state requires prior approval	Not required by Claim 1; optional alert/prompt features appear only in Claims 6-8
Deploy the fix	Scheduled release to production	System generates a production model directly from the updated template (Claim 9)

Every human checkpoint in the middle column is optional in the right one.

A Backtesting Engine Patented Just as Regulators Started Requiring One

The loop the patent claims, continuous predicted-versus-actual monitoring feeding an automated correction, is close to what state insurance regulators have spent three years asking insurers to build. The NAIC’s Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted at the 2023 Fall National Meeting, requires a written AI Systems Program covering, among other things, “validating, testing, and retesting to assess generalization of outputs upon implementation” (NAIC Model Bulletin, adopted December 2023). Twenty-nine jurisdictions now regulate insurer AI use, either through the model bulletin directly or through a state-specific framework in California, Colorado, New York, or Texas, according to the NAIC Model Bulletin Implementation Map as tracked through July 2026.

State Farm’s patent describes a system that performs the same validating-testing-retesting cycle regulators want, built into the rating engine itself rather than bolted on as a compliance exercise, and it has patented that capability as a competitive asset rather than published it as a governance practice. The tension is that “testing and retesting” in a bulletin written around human-reviewed model governance assumes a defined, documented model version sitting between tests. A system whose testing step also rewrites the thing being tested does not produce that version; it produces a moving target.

The bulletin’s third-party oversight language adds another layer specific to this architecture: insurers remain responsible for AI systems built or hosted by outside vendors, a rule actuary.info covered when the NAIC formalized it. Figure 1 of State Farm’s patent wires an external trained LLM and external data sources directly into the same loop as the internally hosted code LLM, meaning any GPT model licensed from an outside vendor and used to satisfy claim 1 inherits the same accountable status as software State Farm wrote in-house. The architecture does not distinguish; the regulation does.

Rate Filing Cadence Against a Same-Cycle Rewrite

The pace mismatch is concrete. Milliman’s regulatory tracking put the countrywide four-quarter rolling average for personal auto rate filing approval at 67 days as of the second quarter of 2025, with California, the slowest jurisdiction it tracks, averaging 234 days (Milliman, Q2 2025). The Insurance Research Council’s most recent look at the same pipeline found the average climbed from 39 days to 54 days between 2010 and 2023, a 40% increase, even as the number of filings insurers withdrew rather than see rejected grew from roughly 1,900 to 3,200 a year (Insurance Business America, citing IRC data). Guidewire’s own rate-filing primer puts the practical distance between the historical data an insurer starts from and the day a new rate reaches a policyholder at up to 18 months once compilation, model build, and regulatory review stack together (Guidewire, 2026).

Claim 10 lets the GPT model train for a specific jurisdiction, meaning nothing in the patent stops a carrier from regenerating its California rating logic on one internal cycle and its Texas logic on another, independently, each time claim 5’s threshold trips. Nothing ties that regeneration cadence to a state’s filing calendar. A rate filing governs the rating factors and relativities an insurer is permitted to charge, not the software architecture that derives them, so a self-updating engine does not violate a prior-approval statute by existing. But a system that can identify a miss, rewrite the code, and validate the rewrite inside a processing cycle measured in hours can produce a materially different rating algorithm well before California’s 234-day median review of the version it replaces even clears the queue. actuary.info has tracked this drift problem emerging across the industry’s pricing AI generally; State Farm’s patent is the first place the mechanism producing it has shown up in claim language.

Run the hypothetical forward. A Texas auto book generates enough predicted-versus-actual drift in a single quarter to trip claim 5’s threshold. Under claim 10, the GPT model retrains on Texas-specific data, the code LLM rewrites the Texas sub-model, the system validates it against Texas loss history in simulation, and claim 9 pushes a new production model, all while the carrier’s most recent Texas rate filing is still sitting in a regulator’s queue behind filings from carriers that never touched their algorithm at all.

Where Model Risk Management Assumes a Version Number the Patent Doesn’t Guarantee

Model risk management, as banks and insurers have practiced it for a decade, validates one specific, frozen model version: a document describing its assumptions, a back-test against a hold-out sample, a sign-off, a version number a rate filing or reserve opinion can cite. Grant Thornton’s 2026 AI Impact Survey, fielded between February 23 and March 18 among insurance executives, found 44% saying governance or compliance challenges had already contributed to an AI project failing or underperforming, before any model in the sample rewrote its own architecture (Grant Thornton, 2026). Fewer than 25% of enterprises across sectors have implemented a structured model risk management framework built for generative AI at all, according to risk-advisory firm Pirani Risk (Pirani Risk, September 2025).

A framework built to validate one static version cannot, by construction, validate a system whose defining feature is that it produces a new version every time its own GPT model flags a threshold breach. actuary.info has documented the same structural mismatch on the claims side, and the exclusion runs deeper than practice: federal bank guidance on model risk management does not extend to insurers at all, a gap this site has covered directly. State Farm’s patent is the first case where that regulatory gap and a live, patented product occupy the same filing.

Who Certifies a Rate the Model Wrote Itself

There is a second, quieter exposure buried in claim 8’s list of triggering data elements: claims history, vehicle history, prior insurance history, and public records. A model that repeatedly regenerates its own rating code against those four inputs, without a defined stopping point, can drift toward a proxy variable, a public record correlated with a protected class, for instance, without ever running through whatever bias review a human-supervised process would have caught. actuary.info has covered this unintended-bias risk for static pricing models; a model that edits its own risk-classification logic on a rolling basis multiplies the number of untested code paths a bias audit would need to catch, without multiplying the audits themselves.

Most states require the actuary certifying a rate filing to attest, in an actuarial memorandum, that the resulting rates are not excessive, inadequate, or unfairly discriminatory, a certification tied to a specific, documented rating algorithm as of the filing date. Claim 1 does not require a human to review, let alone certify, the model software template the code LLM generates before claim 9’s production model goes live. Claims 6 through 8 add a natural-language interface, a detection alert, and a prompt-driven review, but all three are dependent limitations; a system practicing only the independent claim satisfies State Farm’s own patent without a human ever seeing the code change. actuary.info raised the identical accountability question on the reserving side when large language models began drafting reserve opinion language; State Farm’s grant moves the same question from a document an actuary signs to a rating engine an actuary is supposed to be able to explain.

None of this requires new legislation to bite. Existing rate-regulation statutes already hold the filing insurer, and the actuary whose name appears on the supporting memorandum, responsible for the rates charged regardless of what produced them. A state examiner auditing a filing built on this architecture will ask for the version history of the rating algorithm in effect on the date of loss, the same request examiners make of any carrier today. The patent does not describe how State Farm intends to answer that request; it describes a system built to make the question harder to answer with a single number.

Whichever credentialed actuary’s name sits on a rate filing built through this pipeline is underwriting legal exposure for a code change they were never required, by the patent’s own core claim, to review. A documented human checkpoint between the code LLM’s output and claim 9’s production deployment is not something the patent mandates; it is something a carrier will have to add on its own, because no regulator is going to accept that the model updated itself as an answer to the question of who approved the rate.