A Claims AI Patent Now Reaches Into Loss Reserving

Every first-notice-of-loss report forces the same operational fork: tow the wreck to a scrap yard or route it to a repair shop. US Patent 12,694,454 B2, granted to Assured Insurance Technologies on July 28, 2026, wires that fork directly into a named "reserve estimation module" that outputs the dollar case reserve, folding what has always been a claims-workflow decision into a reserving one.

23.1%

Share of all 2025 US auto claims CCC Intelligent Solutions classified as total losses, a record high

Assured patent filings claiming the same named reserve estimation module, one granted, one published

$1B

Assured's unicorn valuation as of April 3, 2025, five months before the classifier application that became this patent

actuary.info has tracked the carrier side of 2026's insurance AI patent wave, including July's finding that USAA took 19 utility grants across the month's three USPTO issue dates while State Farm, the industry's largest AI filer, took none (USPTO Official Gazette, July 2026). US 12,694,454 B2 comes from a different kind of applicant. Assured Insurance Technologies is a claims-processing vendor, not a carrier, and the reserve estimation module its patent describes does not stop at deciding where a wrecked vehicle goes. It decides what the carrier holds on its books for the claim, while the adjuster's file is still open.

What US 12,694,454 B2 Actually Claims

The patent, titled Machine-Learning Method of Loss Prediction Using Incident Information, issued July 28, 2026 from an application filed April 24, 2025, itself a continuation of an earlier application filed August 19, 2024 (US Patent 12,694,454 B2, July 2026). Its abstract states plainly what the system does: incident data tied to a vehicle collision goes in, and a "total loss prediction... indicative of whether the vehicle is repairable or totaled" comes out (US 12,694,454 B2, July 2026).

Claim 1 describes more than a binary call. The system trains its model on historical vehicle incident data using discrepancy analysis, presents a three-dimensional representation of the vehicle so a user can mark damage locations, and generates a prediction of whether the vehicle is "repairable or totaled relative to a value" (US 12,694,454 B2, July 2026). A human still marks the damage on the 3D model in this claim; the model does not work from raw photos alone. What happens next depends on the model's answer. A totaled vehicle triggers coordination of a tow service to a scrap yard; a repairable one triggers selection of a service provider and scheduling of repairs. Claim 6 widens the output to a three-way state, repairable, totaled, or salvageable, recognizing that a heavily damaged vehicle sometimes has more value stripped for parts than sold intact to a body shop.

Model output	Downstream action claimed
Totaled	Coordinate a tow service to a scrap yard
Repairable	Select a service provider; schedule repair
Salvageable (claim 6)	Distinct third state alongside repairable and totaled

The Reserve Estimation Module: Where the Classifier's Output Lands

The classifier's output does not stay inside the workflow layer. The specification names a "reserve estimation module 170" that ingests the full corpus of information a claim accumulates: contextual claim data, damage specifics, injury reports, medical reports, police reports, fraud-detection results, vehicle-collision simulations and guided content capture evidence (US 12,694,454 B2, July 2026). From that corpus, the module generates "a total amount of cost arising from the claim event," covering medical costs, repair costs, item and part replacement, and service-provider costs, using a model trained on historical claim data and real-time cost information.

The training method is the more interesting claim for a reserving actuary. The model processes closed claim files that carry both an initial reserve estimate and the final payout on the claim, identifies discrepancies between the two, and fine-tunes its parameters to shrink that gap on the next prediction (US 12,694,454 B2, July 2026). That is a reserve-adequacy backtest running inside a training loop rather than inside an actuary's development triangle. The estimate itself is not static either: the module "adjusts as the corpus of information corresponding to a particular claim is being built," generating something closer to a running case reserve than a single number fixed at intake.

Assured is not describing this module once. A related application, US 2026/0050991 A1, Trained Machine Learning Model for Optimized Reserve Estimate Prediction, filed April 25, 2025 and published February 19, 2026, claims the same architecture on its own: a system that accumulates completed claim files, trains a model on them, and executes the trained model on new claims to generate optimized reserve estimates (US 2026/0050991 A1, February 2026). Read together, the two filings show a company patenting the reserve estimate as an invention in its own right, not as an incidental byproduct of a claims-triage tool.

Triage Router or Reserving Model? The Line the Patent Blurs

CCC's and Mitchell's estimating platforms have automated total-loss determinations for years, and the industry has never treated that as an actuarial event by itself. A triage tool tells the claims system where to route the file, tow truck or repair shop, in-house adjuster or a delegated-authority vendor. It does not, on its own, claim to set what the carrier holds in reserve. US 12,694,454 B2 collapses that separation by design. The same total-loss prediction that decides the tow dispatch also feeds the module that prices the claim, and the patent names both functions inside a single system rather than describing a workflow tool and a reserving tool that merely happen to share data.

That distinction is not academic to the actuary who eventually signs a reserve opinion. A misrouted tow dispatch is an operational inconvenience, corrected the next day. A misclassified total loss that has already generated a case reserve, and possibly a settlement offer built off it, is a reserving error that has already started to develop before anyone notices it. With CCC putting 2025's total-loss share at a record 23.1% of all US auto claims (CCC Intelligent Solutions, January 2026), and driveable-vehicle total losses more than doubling since 2021, the volume of claims where that distinction is live has never been larger.

Case Reserve, Not IBNR, But the Bias Travels

Every claim the reserve estimation module touches has, by definition, already generated a first notice of loss. That places the module's output squarely in case-reserve territory, the dollar estimate carried against a specific, known claim, rather than incurred-but-not-reported reserves, which cover losses that have happened but have not yet been reported at all. Nothing in the patent's claims reaches into IBNR estimation, and nothing here should be read to suggest that it does.

The two claim types the classifier separates do not develop the same way, and that is where a classification error does more damage than a simple dollar miss. A total-loss claim converges to actual cash value quickly, typically settled within weeks once ownership of the wreck transfers, with little tail beyond a possible valuation dispute. A repairable claim carries genuine development risk: supplemental damage discovered once a panel comes off, rental-car extension while parts back-order, and, in a meaningful share of cases, a mid-repair reclassification to total loss once supplements push the estimate over a state's threshold. Call a claim repairable that should have been totaled, or the reverse, and the case reserve is not merely off by a margin. It is sitting on the wrong development curve entirely, carrying a completion pattern that belongs to a different type of claim. A claim reserved as a $4,000 repair that later re-flags as a $14,000 total loss does not just understate severity by $10,000; it also inherits a settlement timeline, salvage recovery entry and closure pattern the original case reserve never modeled.

That is where a case-level bias stops being a case-level problem. If the classifier systematically miscalls total losses in one direction, whether from training data that undersamples the newer, sensor-dense vehicles now driving CCC's record total-loss share, or from any other source of drift, the error does not net out in the aggregate. It shows up as a shift in the paid-to-case development diagonal, the exact signal reserving actuaries use to select IBNR and IBNER factors off a loss triangle. actuary.info's earlier look at how faster AI-driven claims handling tests IBNR bias in workers compensation found the same structural risk in a different line: speed and automation at the front of the claim can shift the diagonal actuaries rely on to read the back of it. A module that claims only to set a case reserve is, one link downstream, an input to the number every actuary spends far more time on.

A Claims Vendor's Patent on the Actuary's Number

Assured Insurance Technologies is not a carrier. The Palo Alto company, founded in 2019 by Theo Patt and Justin Lewis-Weber, builds first-notice-of-loss, messaging, fraud-detection and catastrophe-response tools that it says process tens of millions of claims a year for the nation's top insurers (Assured, 2026). It reached a $1 billion unicorn valuation as of April 3, 2025, following a March 5, 2025 seed round co-led by Iconiq Capital and Kleiner Perkins with 18 institutional investors participating, and carried 199 employees as of June 30, 2026 (Tracxn, 2026).

That makes US 12,694,454 B2 a different kind of grant than the carrier-held patents actuary.info covered from July's USPTO issue dates, where USAA and Allstate patented systems built on their own claims data and deployed inside their own operations. Assured's patent fences off an architecture it can license or embed across every carrier that buys its claims platform. A carrier adopting Assured's FNOL and claims-processing stack does not just get faster intake. Per the patent's own architecture, it gets a reserve number generated by code the carrier does not own, running on training data whose provenance, pooled across Assured's full carrier customer base or segmented client by client, the patent's public claims do not specify.

Rating factors face a filing requirement precisely because a regulator, and an appointed actuary, need to see the mechanics before the number reaches a policyholder's bill. A case reserve carries no equivalent filing step. It lands directly on the loss and loss-adjustment-expense reserve schedule, reviewed after the fact by the actuary signing the statement of opinion and by auditors, rather than vetted before the fact by a regulator. A vendor's patented, largely opaque classifier feeding that number at the moment of FNOL is a governance gap that rating algorithms, for all their own AI controversies, do not share in the same way. actuary.info's coverage of MassMutual's newly granted mortality-score patent traced a parallel dynamic on the life side, where a patented model output feeds an actuarial number rather than merely informing a business decision around it.

What a Vendor-Origin Reserve Estimate Puts on the Signing Actuary's Desk

ASOP No. 43, Property/Casualty Unpaid Claim Estimates, adopted by the Actuarial Standards Board in June 2007, directs the actuary to evaluate the data underlying an unpaid claim estimate and to account for changes in the claims process that could affect its reliability. The standard predates AI-generated case reserves by nearly two decades, but a carrier swapping adjuster-set case reserves for a vendor's machine-learning output at FNOL is exactly the kind of claims-process change its data-reliability provisions were written to catch, whether or not the vendor markets its product as a reserving tool.

The questions that raises are specific rather than rhetorical. Whose closed claim files trained the discrepancy-correction loop the patent describes, a single carrier's book or a pool spanning Assured's full client roster? How often does the model retrain, and does a carrier see the retraining before it changes the case reserves flowing onto its books? Does the real-time reserve estimate that adjusts as a claim's corpus builds get reconciled against the carrier's own internal reserve-adequacy studies, or does it simply overwrite the prior figure? None of these questions requires assuming bad faith on the vendor's part, and none are unique to Assured. They are the review steps a case reserve generated inside patented, licensed code requires that one set purely by an adjuster's judgment never did. actuary.info's look at CCC and EvolutionIQ's workers compensation claims AI raised a similar audit gap in a different specialty line, where faster claims handling changed development patterns before the underlying model's governance had caught up.

The pattern is likely to repeat. As more claims platforms follow Assured in patenting a named reserve-estimation component, rather than treating the case reserve as an incidental output of a triage tool, a growing share of any carrier's book will carry case reserves whose first draft originated outside the carrier's own actuarial function. The signing actuary's job does not change in kind. It changes in where the audit trail has to start.