Insurance AI Regulation Moves Into Market Conduct Exams

Insurance AI regulation crossed a line this year. In March 2026 the National Association of Insurance Commissioners put an AI Systems Evaluation Tool into a twelve-state pilot, giving regulators a written framework for pulling apart an insurer’s algorithms during a market conduct examination. The same year, several statehouses took up bills that would bar an algorithm from being the only thing standing between a claim and a denial.

Insurance AI Regulation Reaches the Examination Room

For most of the last decade, insurance regulators knew that carriers were running claims and pricing decisions through machine learning systems, and had almost no structured way to inspect them. That changed with a document the NAIC calls the AI Systems Evaluation Tool. The tool was built for regulators to use in a market conduct, financial analysis, or financial examination context, and it gathers information about how widely a company uses AI, how that use is governed, which models the company treats as high risk, and what data feeds those models.

As of March 2026, twelve states are piloting the tool. The Big Data and Artificial Intelligence (H) Working Group is collecting feedback through the pilot period, and the NAIC anticipates the tool will be adopted at the 2026 Fall National Meeting.

The tool did not appear out of nowhere. In December 2023 the NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, which reminded carriers that decisions made or supported by AI must still comply with all applicable insurance laws and regulations, set out expectations for how insurers govern that use, and told companies what a department might request during an investigation or examination. The bulletin was guidance, not a statute. The evaluation tool is what turns that guidance into a set of questions an examiner can actually put in front of a company.

The distinction is practical. A bulletin tells a carrier what good behavior looks like. An examination tool tells an examiner what to ask for, in what order, and what an incomplete answer looks like. The published draft of the evaluation tool is organized around governance structure, model inventory, data integrity, and third-party dependencies.

How Far Automation Has Spread Through Auto Claims

The scale is easy to underestimate. Beginning in 2021 the NAIC surveyed insurers line by line to learn how AI and machine learning were being used and what controls were in place. Out of 193 auto insurers that responded, 88 percent reported that they use, plan to use, or plan to explore AI and machine learning models in their operations. Among 194 home insurers, the figure was 70 percent.

The survey also describes what the models actually do inside a property and casualty carrier. In marketing, common uses included targeted online advertising and offers to existing customers. In underwriting, models handled renewal evaluations and inspections. In pricing, machine learning drove risk scoring and rate factor relativities. And in claims, AI is used for accident image analysis and to estimate ultimate claim settlement values, along with fraud detection.

That last item is the one worth reading twice. A system that estimates the ultimate settlement value of a claim is producing a number that shapes what a person is offered. The NAIC notes that roughly half of the models used for marketing came from third-party vendors, while auto and home insurers mostly developed pricing and underwriting models in house.

• Image Analysis. Photographs of damage are scored by software before a human sees the file.
• Settlement Estimation. Models produce a projected ultimate value that frames the negotiation range.
• Fraud Scoring. Files are flagged by pattern, which can slow payment on a legitimate claim.

The reach is widening. As of June 2026, systems were being designed to review thousands of claim files, estimates, photographs, adjuster notes, communications, and policy documents at once to identify patterns relevant to causation disputes, coverage questions, and damage assessments, and to flag whether a particular claim deviates from typical loss patterns. Litigation forecasting is part of the same push: models weigh historical outcomes, venue tendencies, and claim characteristics to generate probability estimates about the likely result of a lawsuit.

State Lawmakers Tried to Draw a Line

While regulators built an examination framework, legislators went at the problem from a different direction. Several 2026 bills shared one idea: an algorithm can recommend, but a human has to decide.

Florida’s House Bill 527 is the clearest example. The bill would have prohibited insurers, workers’ compensation carriers, and health maintenance organizations from reducing a claim payment, denying a claim, or denying part of a claim based solely on the output of an artificial intelligence system, algorithm, or machine learning system, and would have required that any such decision be made by a qualified human professional. It expressly allowed those systems to be used to generate recommendations, and it required carriers using them to document that use in claims manuals subject to regulatory review. The bill carried a July 1, 2026 effective date and died in Rules on March 13, 2026.

A failed bill still tells you something. The drafting shows what regulators and legislators consider the pressure point, which is the moment a model output becomes a decision rather than an input. That framing is now common across state proposals, and it lines up with the governance expectations in the NAIC bulletin.

Industry Pushback and the Federal Question

The evaluation tool did not arrive uncontested. In December 2025, trade groups representing life, health, property and casualty, mutual, and reinsurance carriers signed a joint letter objecting to the pilot. Their stated concerns were that participation was voluntary for regulators and compulsory for companies, that the pilot had no defined duration, that findings gathered during a development exercise could be used against companies, and that the pilot might begin before the final version of the tool was exposed for comment. The groups asked that participation be voluntary and that information gathered be used only to develop the tool.

Regulators went forward anyway. Iowa Insurance Commissioner Doug Ommen, chairing the working group meeting, described the pilot as instructive and said refinements would follow the pilot period. The same reporting notes that regulators are not required to use the tool; it is one more option available during a market conduct examination.

Sitting behind all of this is a jurisdictional dispute. The same InsuranceNewsNet report describes a December 2025 federal executive order establishing a single national regulatory framework for artificial intelligence, which cuts against the authority of individual states. Ommen’s response was that state commissioners retain authority to supervise AI and that any attempt to block that work would affect consumers negatively. Separately, the National Council of Insurance Legislators has been developing a model act on insurers’ use of artificial intelligence, which had not been finalized as of that reporting.

State Enforcement Over an AI Cancellation

Regulators were still building the examination framework when a separate lane produced a concrete result. In May 2026, Pennsylvania Attorney General Dave Sunday announced an agreement with GEICO over an artificial intelligence initiated auto policy cancellation process the state challenged.

The matter began with a complaint from a new policyholder in West Philadelphia who lost coverage during the company’s standard 60 day review for new customers. An AI tool selected her for further underwriting review. She was required to submit additional documentation under threat of cancellation, believed she had submitted what was asked for, and was never told her submission fell short. The policy was cancelled without her realizing it, and she drove uninsured.

Without admitting any violation, GEICO agreed to follow the Pennsylvania Insurance Department’s guidance on insurer use of artificial intelligence. The company added a week to the submission window for new policyholders under review, agreed to accept one form of residency verification instead of two, agreed to accept a current driver’s license as proof of residency where the address matches the policy, and agreed to retrain customer service representatives.

No new statute was required. The state used ordinary consumer protection authority and reached a carrier that ranks third in the Pennsylvania private passenger auto market. That is the current shape of enforcement in this area: existing law applied to a new technology, one file at a time.

Class Certification in Total-Loss Valuation Cases

The regulatory line and the litigation line moved in opposite directions this spring. On April 24, 2026, the United States Court of Appeals for the Sixth Circuit, sitting en banc, ruled 10 to 7 that roughly 90,000 Tennessee policyholders cannot pursue a collective breach of contract claim against State Farm over a valuation input known as the typical negotiation adjustment. The decision makes the Sixth Circuit the sixth federal appeals court to block class certification in this category of actual cash value dispute, joining the Third, Fourth, Fifth, Seventh, and Ninth Circuits.

The mechanic is worth following, because it shows how a single automated step compounds. When a vehicle was declared a total loss, State Farm used a database maintained by Audatex to generate a valuation report. Audatex identified comparable vehicles listed for sale, pulled their advertised prices, and applied the typical negotiation adjustment, a percentage reduction built on the assumption that used car buyers negotiate below the asking price. It then adjusted for mileage, equipment, and options, averaged the results, and applied a condition adjustment.

In the named plaintiff’s file, the report found four comparable minivans advertised between $15,800 and $18,803. The negotiation adjustment removed between $790 and $940 from each of those figures.

Three Appraisers, Three Numbers

After she sued, State Farm invoked the policy’s appraisal clause. The insurer’s appraiser valued the vehicle below what she had already been paid. Hers came in far higher. The neutral third appraiser landed higher still, and the binding written decision followed that figure, with State Farm paying more than $4,000 on top of the original payment. The same vehicle, run through a different process, produced a materially different answer.

The majority held that even if the adjustment was generally flawed, proving that would not establish that any individual policyholder was underpaid, because the policy required payment of the market value of each specific vehicle. Resolving that question across 90,000 vehicles would require individualized evidence on year, make, model, mileage, options, and pre-loss condition.

The dissent, authored by Judge Gibbons and joined by six colleagues, disagreed. It noted that State Farm had conceded that under its reading of the law, an insurer could defeat class certification even if it calculated actual cash value by throwing darts at a board. The dissent also pointed to a parallel Arkansas case, Chadwick v. State Farm, where a jury tried the same theory against the same adjustment and returned a verdict for the class, and where a proposed class settlement has since received preliminary approval.

The practical result is that a challenge to an automated valuation increasingly has to proceed one file at a time, with vehicle specific evidence, rather than as a group attack on the model itself.

What Automated Review Changes for Injured People

Most of this activity is invisible from the outside. Someone recovering from a collision does not see a model inventory or a governance framework. What they see is a number, a timeline, and a reason.

Automation changes three things about that experience. The first is speed in one direction only. Software can close a straightforward file quickly, which is good, and it can also flag a file for review and slow it down for reasons the claimant never learns. The second is anchoring. When a model produces an estimated ultimate settlement value early, that figure tends to frame every conversation that follows, even after new medical records arrive. The third is consistency of the wrong kind. A human adjuster who misreads one file misreads one file. A model that undervalues a category of injury undervalues every claim in that category until someone notices.

None of that suspends existing law. The NAIC is explicit that when insurers use AI they remain responsible for complying with insurance laws, regulations, and consumer protection rules, including requirements related to accuracy and avoiding unlawful discrimination, and that human oversight remains part of insurance decision-making. Every state has statutory standards governing how claims are settled, which we cover in our post on claims and settlement practices. An automated recommendation is not a defense to a claims-handling violation.

It also does not change how liability gets established. The evidence that decides a disputed claim is still the evidence: the physical proof, the medical records, the reconstruction. We have written before about how insurance companies investigate fault, and the underlying method has not changed because a model now scores the photographs first.

Pricing Pressure Behind the Automation

Carriers did not adopt these systems in a vacuum. Auto insurance pricing moved sharply over the last several years, and cost control is the stated reason behind a great deal of claims automation.

The most recent federal data shows the trend cooling. In June 2026, the motor vehicle insurance index declined 2.0 percent over the month. A single monthly decline follows several years of steep increases, and it does not undo them.

Carriers are also in a strong financial position. In the first quarter of 2026, the United States property and casualty industry posted its biggest first quarter underwriting profit in 25 years. Read alongside the adoption figures, the picture is an industry that became faster and more profitable at processing claims at the same moment regulators began asking how those decisions get made.

Where the Two Trends Meet

Pricing discipline and claims automation are the same project viewed from two ends. A carrier that cannot raise rates without regulatory friction looks for savings on the payout side, and software that estimates settlement values is one of the levers available. That is precisely why an examination framework aimed at claims models arrived when it did.

This is also why the NAIC survey results carry weight beyond their percentages. When 88 percent of responding auto carriers are using or exploring these systems, the question is no longer whether automated evaluation is normal. It is whether anyone outside the company can see how it works.

Documentation in an Automated Review

If a model is reading the file, the file is the argument. That has always been true, and automated review sharpens it.

Structured, complete medical documentation reads cleanly to both a human adjuster and a model. Gaps in treatment, inconsistent descriptions of symptoms, and missing records read as noise, and noise is what pattern-matching systems penalize. Photographs, repair estimates, and contemporaneous records are the inputs the software was trained to weigh.

The practical consequence is that a well-built file is worth more than a well-argued phone call. Our team builds records with the understanding that they will be read by software before a person ever picks up the phone, and we handle the carrier directly so our clients are not the ones supplying an automated system with an off-the-cuff description of their injuries. Our Bigger Share Guarantee® and $0 Out-Of-Pocket Forever structure mean the work happens without cost to the client while a claim is pending. For a closer look at how these systems evaluate individual claims, see our post on AI insurance claim evaluations.