Thirty-one granted US AI patents reached tracked insurance-sector filers in September 2026, counted by grant date from official USPTO records. That is up from 21 in September 2025 and the highest monthly figure of the year. The concentration is the story: nine companies account for the whole month, UnitedHealth and its Optum units took 17 of the 31, and health filers took 19. A month that looks like acceleration sits inside a trailing year that is slightly smaller than the one before it.
Key Takeaways
- 31 grants in September 2026 against 21 in September 2025, a gain of 10, spread across five grant Tuesdays that ran 5, 6, 9, 2 and 9.
- Nine filers appeared in the month, down from 10 a year earlier, with UnitedHealth at 17 and State Farm at 5 taking 71% between them.
- 285 grants in the twelve months to September 30, against 299 in the prior twelve; January through September 2026 ran 224 against 228 for the same months of 2025.
- 39 months was the median filing-to-grant interval for the month's cohort, against 34 months for September 2025; only 3 of the 31 were filed in 2025 or later.
- Six grants protect machinery that evaluates, explains or re-tunes a model rather than machinery that scores a risk.
The Month in Grants, and the Year It Sits In
September's 31 grants beat August's 26 and July's 18, and they beat the 21 the tracker recorded in September 2025. Read on its own the month reads as a step up. Read against the trailing twelve months it does not move the level: 285 grants in the year to September 30 against 299 in the twelve months before, and 224 in the first nine months of 2026 against 228 in the same stretch of 2025. One heavy month inside a flat year is what a lumpy grant calendar looks like, and two of the five issue dates in September carried nine grants each while a third carried two.
Composition moved more than volume. Health filers took 19 of the 31, or 61%, against 6 of 21 in September 2025. Carriers took 10 and data or analytics vendors took 2. The trailing-year picture carries the same rotation: health at 93 grants against 74 a year earlier, carriers down to 132 from 160, vendors to 45 from 50. The property-casualty side of the tracker is issuing fewer AI patents than it was a year ago while the payer side issues more.
Pendency is the figure to hold onto. The median September 2026 grant waited 39 months from filing, against 34 months for the September 2025 cohort, and only three of the 31 were filed in 2025 or later. The slowest in the month was USAA's virtual coaching patent, US 12,731,157, which took 1,988 days from a March 2021 filing. What issued in September is a read on what insurers were building in 2022 and 2023, which matters when the question in front of a pricing or model-risk committee concerns generative and agentic systems that mostly entered the queue after that.
Seven Grants From the Month
| Filer | Patent | Granted | Filed | Subject |
|---|---|---|---|---|
| Allstate | US 12,725,462 | Sept 1, 2026 | Apr 15, 2024 | Suppressing false collision alerts in telematics |
| Allstate | US 12,731,032 | Sept 8, 2026 | Jan 6, 2025 | Retraining a driver-analysis model after deployment |
| UnitedHealth | US 12,731,007 | Sept 8, 2026 | Jun 8, 2023 | Causal transformer selecting actions on encounter data |
| State Farm | US 12,737,817 | Sept 15, 2026 | May 22, 2023 | Home score from inferred component ages |
| Equifax | US 12,737,686 | Sept 15, 2026 | Jun 14, 2023 | Trend attributes as first-class model inputs |
| Cotality | US 12,743,487 | Sept 22, 2026 | Nov 11, 2022 | Neighborhood-specific property valuation models |
| Optum | US 12,749,017 | Sept 29, 2026 | Dec 8, 2022 | Unitless dissimilarity metric for feature bias |
Allstate's two grants bracket the telematics problem from both ends. US 12,725,462 runs machine learning over vehicle and telematics data to produce a collision output, then suppresses it twice before affirming: once by testing whether the reading came from within a set radius of a known false-positive location, and again by scoring the event against a telematics threshold. The output feeds first notice of loss and automated claims triage, so the thing being protected is cleaner loss data. US 12,731,032 goes after the model itself. It clusters pattern deviation outputs so that inter-cluster variance over intra-cluster variance is maximized, which groups model failure modes rather than drivers, trains a long short term memory network per cluster to verify consistency, and feeds only verified clusters back into the analysis model. A driver risk metric that keeps adjusting after deployment is an input to usage-based pricing that moves without a filing.
UnitedHealth's US 12,731,007 tokenizes member encounter records supplied in tuple form, including tokens for the actions taken, and trains a causal transformer to predict the outcome of candidate actions and then to select among them. The output is a recommended action rather than a probability, aimed at care management and utilization decisioning over longitudinal histories. State Farm's US 12,737,817 estimates the age of structural components from home telematics, rolls those ages into a home health indicator standing in for effective age, and converts it to a single home score. Its sibling US 12,737,932, granted the same day, wraps a conversational interface around a water sensor placement model, the third member of the family behind State Farm's placement patent granted in July.
Two vendor grants sit on the data layer underneath pricing. Equifax's US 12,737,686 generates a trend attribute from each entity's time series, in some embodiments by applying a frequency transform and keeping a subset of the coefficients, writes it back into the training structure, and trains on the enlarged feature set. Two files sharing a current value and heading in opposite directions stop looking alike. Cotality's US 12,743,487 clusters properties into contiguous neighborhoods partly from how often a pair is cited as comparable in appraisal reports, then fits a machine learning model inside each neighborhood, which bears on replacement cost and exposure work. Optum's US 12,749,017 reduces a model's per-class performance metrics for one evaluation feature into a unitless dissimilarity metric, so bias results compare across features and across models; a companion, US 12,749,016, filed the same day in December 2022, issued alongside it.
Movers, and What the Month Chose to Protect
Against the prior twelve months, UnitedHealth is up 11 to 73 and holds the leaderboard, while State Farm is down 15 to 61 and Allstate down 8 to 28. Humana is up 7 to 10, Cotality up 6 to 8, CVS Health and Elevance up 4 each to 5. AIG moved from zero to 8 over the year, though none of those landed in September. Going the other way, MassMutual is down 7 to 5, Verisk down 6 to 8, Liberty Mutual down 5 to 1 and Travelers down 4 to 6. Thirty-two tracked companies recorded at least one grant in the trailing year; 145 recorded none, which is the more useful number when a vendor claims a patented capability.
The theme in September is governance machinery. Six of the 31 grants cover tooling that evaluates, explains or re-tunes a model rather than tooling that scores a risk: the two Optum feature-bias patents, Optum's US 12,749,020 on moving a confidence threshold by reinforcement learning against a reward signal, Optum's US 12,743,621 on inspecting how a model trained through interpolated Shapley values, Optum's US 12,737,645 on scoring the quality of a graph training dataset, and Allstate's on-deployment retraining loop. Five of the six carry an Optum entity as the applicant. The confidence-threshold patent is the clearest about why: in health operations the threshold decides which cases a model disposes of and which route to a human reviewer, so it sets the automation rate and the review cost directly.
That has an awkward edge for anyone relying on documented model governance. A threshold that re-tunes itself against a reward, and a driver model that modifies itself after deployment, both describe systems whose behavior on the day differs from their behavior at validation. The bias-audit regimes now arriving at state level assume a model that can be tested and then held still. These filings protect the opposite property, and they were drafted in 2022 and 2023, before most of those rules existed. The monthly series is tracked at AI Patent Watch, with per-filer views and the published-application pipeline.
Further Reading
- Insurer AI Patent Grants, July 2026 – the prior monthly read, when carriers rather than payers led the count.
- State Farm, USAA and Allstate Hold the Carrier AI Patent Concentration – the carrier side that is now shrinking year over year.
- Humana's Time Series Model Selection Patent and Trend Work – the September 1 Humana grant, read in full.
- Federal Circuit Section 101 Rulings and Insurer AI Patent Moats – how much of a granted claim survives an eligibility challenge.
- Insurance AI Patents Are Narrowing in Prosecution – why a 39-month pendency produces a thinner claim than the one filed.