NAIC Targets AI in Claims Handling at Spring 2026 Meeting: What Reserving Actuaries Need to Prepare

At the NAIC's Spring 2026 National Meeting on March 24, consumer representatives on the Big Data and Artificial Intelligence (H) Working Group flagged AI in claims handling as an area needing further review, citing transparency gaps in coverage determinations and the potential for model bias to produce outcomes that would breach state unfair claims settlement practices acts. It is the first time the Working Group has separated claims from pricing and underwriting.

Key Takeaways

88% of 193 responding auto insurers use, plan to use, or plan to explore AI and machine learning, on the NAIC's own survey data.
Total-loss frequency reached 22.8% through October 2025 while repairable claims fell 10.4% through August, so AI is influencing the repair-versus-total-loss call on roughly one in five claims.
Full-scale AI adoption went from 8% to 34% between 2024 and 2025, but only 22% of carriers that tested AI reached full production deployment.
The vendor registry framework was narrowed to pricing and underwriting only, deferring claims, utilization review, marketing, and fraud detection from its first iteration.
Arizona takes effect July 1, 2026 and Colorado June 30, ahead of the evaluation tool's expected adoption in November.

Why Claims Got Separated Out

The distinction the consumer representatives drew is a substantive one rather than a procedural one. Prior sessions, going back to the 2023 Model Bulletin and the 2025 request for information on a possible AI model law, treated AI across insurance operations as a single governance problem. The argument at Spring 2026 was that the harm from an algorithmically biased coverage denial is immediate and identifiable in a way that an opaque pricing factor is not.

The penetration data supports the timing. Among 193 responding auto insurers, 88% reported using, planning to use, or planning to explore AI and machine learning. Photo-based damage estimation, automated triage scoring, total-loss threshold determination, fraud detection, and subrogation identification are standard components of a mid-size or large carrier's claims stack.

The trajectory matters more than the level. Full-scale adoption expanded from 8% to 34% between 2024 and 2025, while only 22% of carriers that tested AI reached full production. That gap is itself the governance concern: a pilot running inside claims does not carry the documentation infrastructure a production system requires, and the pilots outnumber the production deployments.

Meanwhile the decision AI is influencing has been moving. Repairable claims fell 10.4% through August 2025 against the prior year, and total-loss frequency reached 22.8% through October 2025, on pace for a second consecutive record. When a model is making or shaping the repairable-versus-total-loss determination for roughly one claim in five, and that ratio is shifting toward total losses, a regulator has a concrete consumer-protection basis for asking how the determination is made.

Where Claims AI Reaches the Reserve

The AI Systems Evaluation Tool, piloting across 12 states from March through September 2026, was not built for claims, but its exhibits reach them. Exhibit A asks carriers to count production models by use case and consumer impact, which for claims means every scoring model in the workflow including ones embedded in vendor platforms that adjusters use without thinking of them as AI. Exhibit C asks for development documentation, testing evidence, and human-in-the-loop protocols on high-risk models.

The actuarial exposure sits downstream of both. Claim severity estimates, closure rates, and settlement amounts influenced by AI flow into the data that sets case reserves and estimates IBNR. That is not a question of validating the vendor's model; it is a question of whether the loss development triangle reflects a settlement process that changed.

It does change. An AI triage system that routes straightforward claims to automated settlement while flagging complex claims for manual review produces a different development pattern than the one generated when adjusters handled both. Factors selected from pre-deployment experience and applied to post-deployment periods will misstate ultimates, and factors fitted on AI-influenced experience will not transfer to states or lines where the system is not running. Dating the operational change is the ordinary part; documenting which diagonal it lands on with reference to the model is what the exhibits will ask for.

The override rate is the metric that carries the most weight, and few carriers publish one. Model validation practice in rate filings already documents how recommendations are accepted, modified, or overridden. The claims analogue is how often an adjuster moves an AI-recommended reserve, in which direction, and whether that pattern varies by claim type or claimant profile. A model that tests cleanly can still produce a biased outcome through an override pattern that does not.

There is a scope asymmetry working against carriers here. At the same meeting, the Third-Party Data and Models Working Group narrowed its vendor registration framework to pricing and underwriting, deferring claims. So the evaluation tool asks a carrier about vendor oversight and data lineage for claims models built by CCC, Tractable, CLARA, or Verisk, while those vendors sit outside the registry that would give regulators direct visibility. Contractual audit rights and bias testing evidence have to come from the carrier's own agreements.

The States Are Moving Faster and More Prescriptively

The NAIC's instrument is a governance questionnaire. Several states have gone straight to mandates, and their dates land before the tool is finalized.

State	Bill / Regulation	Effective Date	Claims-Specific Requirement
Florida	HB 527	Died March 13, 2026	Qualified human professional for all claim denials; detailed recordkeeping
Arizona	HB 2175	July 1, 2026	Licensed medical director must personally review and sign health claim denials
Colorado	SB 24-205	June 30, 2026	Algorithmic discrimination testing for high-risk AI affecting coverage decisions
California	AB 3030	Enacted	Transparency disclaimers for AI-generated clinical information in coverage decisions
Texas	HB 149 / SB 1188	Enacted	Practitioner disclosure of AI use in diagnostic recommendations

Florida HB 527 is the template even though it failed. It would have barred workers' compensation carriers, insurers, and HMOs from using AI as the sole basis for denying or reducing a claim, required a qualified human professional to independently analyze the facts and review the AI output, and mandated records of that person's identity, the timing, and the basis for the determination. It passed the House and died in Senate Rules on March 13, 2026.

Colorado's approach is different in kind. SB 24-205 requires documented testing for algorithmic bias in high-risk systems affecting coverage decisions rather than mandating a human reviewer, which is a heavier evidentiary burden than the NAIC Model Bulletin's general principles and a lighter operational one than Florida's.

NCOIL sits between them. Its model act, sponsored by Assemblyman Erik Dilan and Representative Forrest Bennett, would require a qualified human professional to make the final decision on all claims, reviewing AI outputs for accuracy and prior human decisions on the claim, with records of every action. ACLI, NAMIC, APCIA, and AHIP opposed it jointly as a one-size-fits-all approach duplicating technology-neutral law already on the books, and development was paused.

NCOIL resumes at its November 2026 annual meeting, which is when the NAIC expects to finalize the evaluation tool. A carrier could therefore face a prescriptive human-in-the-loop mandate and a documentation-based governance framework emerging in the same month, and a multi-state book has to satisfy the stricter of the two wherever either is adopted. The risk classification a carrier chooses determines which of its claims models fall inside that overlap, and a methodology built when regulatory attention sat on pricing models will not put them there.

NAIC Targets AI in Claims Handling at Spring 2026 Meeting: What Reserving Actuaries Need to Prepare

Key Takeaways

Why Claims Got Separated Out

Where Claims AI Reaches the Reserve

The States Are Moving Faster and More Prescriptively

Further Reading

Sources