Allstate's Patent Puts a GPT Between the FNOL Call and the Payout Model

Allstate was granted US Patent 12,705,674 B2 on August 11, 2026 (USPTO Official Gazette) for a claims pipeline in which a generative model reads a policyholder's free-text account of an accident, extracts more than twenty structured claim variables, and keeps generating follow-up questions until the required set is complete. Those variables then go to a model correlating them to claim outcome, including payout amount. The language model is not answering the policyholder. It is manufacturing the covariates a downstream severity model consumes.

Key Takeaways

One of 136 insurer AI patents at Allstate alone, part of a three-carrier bloc holding 77% of all insurer AI filings since 2014.
The required variable set includes no-fault status and distracted-driving statute applicability. A system extracting those is not collecting contact details; it is assembling the record on which coverage and liability get decided.
$41.1 billion of reserves for claims and claims adjustment expense at December 31, 2025 (Form 10-K). Auto claims are the bulk of the count, and initial case reserves are set off exactly the intake information this patent structures.
25.95 million auto policies in force (Allstate, August 6, 2026), so a pipeline that removes handling steps before an adjuster opens a file moves both allocated and unallocated loss adjustment expense at once.

Patent Details

Patent number	US 12,705,674 B2
Title	Claims processing with extracted claim variables
Assignee	Allstate Insurance Company, Northbrook, IL
Inventors	Olivia Mahler-Haug, Corey Fyock, Isabella DeRosa, Nicole Bauer, Julio Castaneda, Daniel Kashkett
Application	18/456,336, filed August 25, 2023
Prior publication	US 2025/0069148 A1, February 27, 2025
Granted	August 11, 2026 (Official Gazette Vol. 1549, No. 2)
Classification	CPC G06Q 40/08, insurance data processing
Claims	19 (system, method, and computing-system sets)

What the Model Is Told to Extract

The specification does not describe a general-purpose summarizer. It names a predetermined required set of claim variables the orchestration circuit must fill before a claim can move, and the list reads like an adjuster's opening script rendered as a schema: tow involvement and body type, door operability, fluid leaks, gross damage, injury occurrence and detail, police report, prior damage, location, no-fault status, phone restrictions, coverage status and liability determination.

Group those and three actuarial families fall out. Tow, door operability, fluid leaks, body type and gross damage are physical-damage severity proxies, separating a drivable fender claim from a probable total loss. Injury occurrence and detail open the bodily-injury coverage, where distributions are heavier and development longer. The jurisdictional variables are liability-determination inputs, and their presence marks this as something other than intake convenience.

Claim 1 recites the loop rather than the extraction. The system receives a first statement, parses it, compares extracted variables against the required set, identifies which are missing, generates a request targeting those gaps, and transmits it. Claim 19 requires an indication that the combined set is complete before anything reaches the claims processing server.

That loop solves a problem actuaries usually meet on the back end as a nuisance in a modeling dataset. Missing fields in historical FNOL data are almost never missing at random: adjusters ask more probing questions when a claim smells expensive, so injury detail and prior-damage history tend to be populated on exactly the files where severity runs high. Any early-severity model fit on that data is learning partly from the fact of a field being answered rather than from its value.

A system that enforces the same required set on every claim changes the missingness mechanism itself. That is a covariate shift, not a data-quality improvement. A severity model validated on pre-adoption records does not automatically hold on post-adoption ones, because the population of populated fields is no longer the same population.

From Free Text to Payout Estimate

The extracted variables go to a claims processing server, an outcome comes back, and the specification says what generates it: a statistical, mathematical or machine learning model correlating the variables to payout amount, vehicle damage severity, total loss determination and at-fault party identification.

Read the architecture carefully and the language model is not the severity model. The GPT produces regressors; a separate correlation model consumes them. That split divides validation responsibility in a way most model risk frameworks are not built to handle. Extraction accuracy is a natural-language problem owned by a data science team and measured in field-level precision and recall. The outcome model is a statistical problem owned, in a well-run shop, by actuaries. Neither team owns the interaction.

The interaction is the classic errors-in-variables problem, and it does not behave intuitively. Measurement error in an explanatory variable attenuates that variable's fitted coefficient and pushes the lost signal into the residual. A correlation model calibrated on hand-coded historical variables and then deployed on machine-extracted ones will systematically under-respond to precisely the variable the extractor reads least reliably.

If door operability is extracted correctly 96% of the time and total-loss propensity leans on it, the deployed model does not fail loudly on 4% of claims. It runs slightly flat on all of them, and the residual variance it exports looks like ordinary process noise rather than a calibration defect. Field-level accuracy reported on its own will not reveal that. What reveals it is a back-test of predicted against realized payout, stratified by whether the variable was machine-extracted or human-entered, which is a test the actuarial side has to ask for.

The Drift That Arrives Without a Memo

If a model-derived severity estimate seeds or influences the opening case reserve, the level of case reserve adequacy at early maturities moves. That is the specific condition the standard correction was written for: Berquist and Sherman restates historical case reserves to the current adequacy level before development factors are selected, precisely because incurred factors drawn from years with different case reserve strength do not describe the current year.

The difficulty is that the adjustment presumes the change can be identified. A claims department revising its case reserving standards issues a memo with a date on it, and the reserving actuary cuts the triangle around that date. An extraction model that retrains and drifts toward slightly different variable values quarter by quarter produces a continuous, undocumented change in the same quantity.

There is no memo, no effective date, and no step change in the diagonal to catch the eye. The adequacy shift appears as a mild, persistent bias in incurred development that a standard chain-ladder selection absorbs into the factors and carries forward.

The synthetic-data circuit makes that worse in the tail. Claims 3, 4, 9, 10 and 15 add a generator producing new claim statements by modifying originals, with the extraction models trained on the synthetic set, on the stated rationale that some scenarios "will not be frequent enough" for effective training.

Rare scenarios in a personal auto book are where the severity tail lives. The model becomes most confident in exactly the region where the book has the least real experience, and that confidence propagates into the payout correlation as narrower predicted variance on tail claims. A stable extraction-accuracy metric is not evidence of stable severity estimation on emerging loss types.

The same identification problem reaches the expense line. Those damage fields decide whether a vehicle goes to a shop, gets a desk appraisal, or is flagged a probable total loss, so capturing them at first contact removes handling steps before an adjuster opens the file.

That compresses allocated expense on the routing side and unallocated on the staffing side at once. Both workhorse ULAE methods assume a stable relationship between claims department cost and either paid loss or claim count, a stability the site examined against Lemonade's 5% LAE ratio, and this pipeline breaks it gradually rather than at a datable boundary.

The NAIC model bulletin names claim administration and payment explicitly, and expects programs to cover ongoing monitoring, updating and retirement. A synthetic-retraining circuit that revises the extractor against realized outcomes is an updating event on a recurring schedule rather than a one-time deployment, which is precisely the case the bulletin's life-cycle language reaches and the one a governance calendar built around annual model reviews will miss.