Fully Automated Claims Decisions Draw Regulatory Pushback as States Rewrite the Rules

The NAIC's Unfair Claims Settlement Practices Model Act requires insurers to adopt reasonable standards for prompt investigation and settlement, to refrain from refusing claims without a reasonable investigation, and to provide a reasonable and accurate explanation for any denial. It was drafted decades before a claim passed through a neural network.

Straight-through processing has meanwhile gone from 10% to 15% of simple personal auto claims to 70% to 90% at AI-deployed carriers. The standard did not move.

Key Takeaways

• 70% to 90% STP on simple personal auto at deployed carriers, against a 10% to 15% baseline, with Travelers reporting 50% or better on its agentic voice assistant in its first quarter.
• 24 jurisdictions have adopted the NAIC's December 2023 Model Bulletin, which clarifies that existing claims practice standards apply to AI-assisted decisions rather than creating new ones.
• Three legislative models have emerged in 2026: mandatory human review, sector-specific prohibition, and comprehensive disclosure with a right to human review.
• January 1, 2027 is when Colorado's SB 26-189 takes effect, with a 30-day post-adverse-outcome disclosure requirement and a consumer right to meaningful human review.
• 68% of carriers outsource AI development while 18% actively track vendor risk, per an AM Best survey of 150 carriers.

The Standard Assumes a Person

Section 4 of Model Law 900-1 prohibits failing to adopt reasonable investigation and settlement standards, refusing to pay without a reasonable investigation, and failing to provide a reasonable and accurate explanation for a denial or compromise offer.

Each of those assumes someone who reads documents, interviews parties, inspects damage and exercises judgment. When a system resolves a claim in seconds, what counts as investigation is an open question, and when it denies one on pattern matching against millions of historical claims, the explanation requirement is the harder half.

That is where the architecture creates the exposure. In a human denial the adjuster's file records what was reviewed, what the policy requires and why the facts fall short, and that record is the defense in bad faith litigation. An algorithmic denial needs an equivalent trace: what data went in, what rules or weights applied, and the reasoning path to the outcome. Deep learning systems used for damage assessment and fraud scoring output a probability or a recommended action, and the path there runs through layers that resist plain-language reconstruction.

Regulators are not asking carriers to stop. A legal analysis published in May 2026 warned that fully automated claims decision-making may violate state unfair claims settlement practices acts, which is a statement about evidence rather than about technology.

Three Statutory Answers to the Same Question

State legislatures have converged on the question and diverged on the answer, and the three approaches impose very different operating costs.

Florida's HB 527 took the strictest line, barring carriers, insurers and HMOs from reducing or denying a claim based solely on the output of an AI system, algorithm or machine learning system, and requiring a qualified human professional to make the decision independently and document how AI was used. It died in the Rules Committee in March 2026, but the distinction it drew, AI-assisted processes permitted and AI-determined outcomes prohibited, is the one other legislatures are reaching for. An amendment had already expanded it from property and casualty to workers' compensation and HMOs.

Nebraska and Georgia took the sector route. Nebraska bars utilization review agents from basing coverage decisions solely on an AI-based algorithm and requires disclosure to the department, providers, enrollees and the public. Georgia's SB 444 goes further, prohibiting an AI system from issuing an adverse determination until a qualified person conducts review with clinical peer participation, and stating that AI systems shall not supersede the clinical peer's judgment. Both are health statutes, and both establish the principle that an algorithmic output cannot substitute for professional judgment in a coverage decision.

Colorado's SB 26-189, signed May 14, 2026 and effective January 1, 2027, does not prohibit anything. It requires pre-interaction notice, a plain-language description of the technology's role within 30 days of an adverse outcome, a right to meaningful human review and reconsideration, a right to correct the personal data used, and three years of record retention. Enforcement sits with the Attorney General with a 60-day cure period expiring January 1, 2030.

The disclosure model is the one that reaches operations without banning anything. Every automated denial in Colorado triggers the 30-day explanation and the review right, so a carrier running an unexplainable model has two options: build an explainability layer, or route Colorado claims through human review. The second choice surrenders the STP gain in that state, which means the compliance cost is measured in the throughput the model was bought to produce.

The Architectures That Produce the Throughput Resist the Record

The tension is structural rather than a matter of implementation quality, and it runs the opposite way from the deployment incentive.

The systems delivering 70% or better STP and the associated cycle time compression are less interpretable than the rule-based engines they replaced. Explainability and throughput are being traded against each other at the architecture level, so a carrier optimizing for one is moving away from the other, and the regulatory requirement lands on the side that automation has been moving away from.

Ownership of the problem is the second complication. An AM Best survey of 150 carriers found 68% outsource AI development while 18% actively track vendor risk, and the claims AI in question frequently comes from third parties. The NAIC's Third-Party Data and Models Working Group revised its framework at the Spring Meeting to limit initial scope to pricing and underwriting, so claims vendors are outside it for now, but the registration mechanism giving regulators direct access to third-party vendors is already built. Extension to claims is a scope decision rather than a new construction.

Examination reach has already moved. The BDAI Working Group's AI Systems Evaluation Tool pilot launched in March 2026 across 11 states, and counsel summaries of the Spring Meeting record it being deployed through market conduct examinations, financial examinations and regulatory inquiries. Claims-handling AI is therefore in scope in examinations that were never triggered by an AI complaint, which removes the assumption that documentation only has to exist once someone objects.