← All mobility signals
Governance

The Audited Write-Off

US regulators are piloting an exam tool that audits insurers’ AI claims models inside market-conduct and financial exams. The collision-repair trade calls it the first exam of total-loss models.

What is it?

Claims models are becoming exam material. The NAIC’s AI evaluation tool asks insurers to count, explain and justify their AI, from claim triage to loss estimation and image analysis, and pilot states now use it inside ordinary market-conduct and financial exams.

Why now?

The pilot began in March 2026 after the working group settled definitions in February. Regulators kept the scope broad, despite an insurer trade body wanting predictive models and GLMs excluded. Trade press reports 12 states in the pilot, with a vote on adoption targeted for November.

Who’s building it?

State insurance regulators and the NAIC, with insurers, their model vendors and consumer representatives arguing scope.

NAICPilot statesAuto insurersModel vendorsConsumer representatives
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — insurers and vendors whose models value claims.
Buyers — drivers whose damaged cars are valued, repaired or written off with a model’s help.
Gatekeepers — state insurance examiners using the NAIC’s evaluation tool in exams.
What’s at stake
Whether a claims model answers to the same unfair claims settlement law that binds a human adjuster.
Vendor models, now named and questioned in the exam.
Which systems an insurer itself classes as high-risk.
What’s unsettled
Where augmenting an adjuster ends and automating the decision begins.
Whether predictive models and GLMs stay inside the scope.
Whether the tool is adopted, and how many states then use it.
02 · How the demand matureda progression, not a timeline
Novelty
Regulators ask insurers to count and describe their AI systems, model by model.
Retention
The questions ride along with exams that recur, not one-off inquiries.
Willingness to pay
Insurers take on documentation, validation and vendor oversight costs to answer an examiner’s questions.
Category
Claims AI becomes a supervised class, with inventories, risk ratings and named vendors.
03 · Where the market goesthe forks, not a prediction
→Inventory or inspection: the tool stays a questionnaire that collects counts and policies, or examiners use it to open specific claims models.
→Adoption in November turns the pilot into a standard exam module across many states.
→Scope narrows to exclude GLMs and older predictive models, leaving only newer AI in view.
→Disputed write-offs and settlements start citing exam findings as evidence against the insurer.
Much depends on who decides what counts as high-risk, since the tool leaves that classification to the insurer whose models are under review.
04 · The Insurer Model deskadjacent market, moving now
The Contested Roof Score
Aerial AI meets a home-insurance bulletin.
Satellite imagery cannot be the sole basis for cancellation.
Kentucky Department of Insurance
The Explained Denial
Automated insurance decisions must now come with reasons.
Colorado’s SB26-189: notice, a reason within 30 days, human review.
Colorado General Assembly
The Liability Question
Liability for AI decisions splits by fault.
Colorado’s 2026 rewrite splits AI discrimination liability by relative fault.
Littler
05 · The Wire
01
Pilot states use the tool in exams.
NAIC Big Data and AI (H) Working Group ↗
02
The tool lists claims uses from triage to imaging.
NAIC ↗
03
Trade press frames the pilot as the first exam of total-loss AI
Autobody News ↗
04
Insurer groups ask regulators to remove pricing models from AI exam tool
Digital Insurance ↗
The panel on this signal
They agreed on
–Claims models, from triage to loss estimation and image analysis, now fall inside the exams insurers already face.
–Regulators kept the scope broad for the pilot.
–Vendor-built models are the insurer’s to explain, with the vendor named in the answers.
The insurance market-conduct examiner disagreed on
–An exam reaches only what the insurer flags: if the company rates its total-loss model low-risk, the detailed questions never apply.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.