Part 4 of 4 in a series on AI and claim denials. Part 1 covered health insurance's sole-basis bans; Part 2 covered where P&C regulation and adjuster licensing are headed; Part 3 covered the litigation already testing automated claims decisions. This part covers what "sole basis" has to mean operationally when deploying a human in the loop, three habits claims orgs can build now, and a closing question about where human review actually helps.
"Sole basis" has no operational definition yet
The concept of sole basis was discussed in prior articles as a guardrail around automated decisions, particularly denials. Yet none of the bills reviewed for this series defines "sole basis" in operational terms, and that ambiguity will matter more as LLMs get better at reasoning and judgment. Florida's HB 527, which passed the House in 2026 and died in the Senate, would have required that, before denying or reducing a claim, a qualified human professional analyze the claim facts and policy terms independently of any AI system and make the determination. Utah's SB 319, enacted in 2026 and effective January 1, 2027, requires that an adverse preauthorization determination on clinical or medical necessity be made by a person who exercises independent medical judgment (unless a specialist is consulted) and does not rely solely on recommendations from any other source. Neither answers the hard question: what separates genuine human review from a human clicking approve on an AI-generated denial in under a minute?
The Florida bill died in Senate Rules; the record does not say why. “Sole basis” could arguably mean that humans would be forever put into the position of reading everything manually, which it's clear they probably won't do and is simply not practical in light of the formidable power of LLMs.
None of the bills reviewed for this series sets a minimum review time or a defined depth of file re-analysis. Florida's HB 527 came closest, listing steps a reviewer must take and records a carrier must keep, but it did not define what counts as sufficient review. A defensible record of human review has to show what the reviewer actually read, what they changed or confirmed from the AI's output, and how their conclusion connects to that file's specific facts. A digital signature with no supporting record behind it isn't meaningfully different from letting the AI decide.
The litigation covered in Part 3 and this documentation question point to the same underlying exposure. In both cases, the exposure isn't the algorithm. It's the absence of a documented, individualized human judgment standing behind the algorithm's output.
What should claims orgs do in deploying AI solutions?
Three habits are worth building now, rather than waiting for a statute to define them:
- Document human review. Record what the reviewer checked on a file before a denial or other decision goes out: what parts of the claim file they read, what they confirmed or overrode from the AI's recommendation, and how long they spent on it. Treating a denial as commodity output — whether the commoditizing is done by an algorithm or a rubber-stamping human — runs against the direction of current regulation. Unfortunately, AI use in claims is often driven by individual adjusters' prompts, making record-keeping around the questions they asked and the output provided difficult.
- Verify authenticity, not just plausibility. As claims orgs lean more on AI to review submitted material — photos, repair estimates, medical records — they increasingly need the ability to tell whether that material is what it claims to be. None of the laws covered in this series requires it. But a human reviewer signing off on an AI-assisted decision needs to be checking two things, not one: whether the AI's read of the material was reasonable and correct, and whether the material itself is authentic and accurate.
- Vet "AI-native" vendors. As more of this work runs through third-party tools, confirm that any "AI-native" vendor actually carries the protections a carrier needs before using its output to support a claim decision: a documented human-review workflow, not just a model that generates a recommendation; an audit trail showing what the AI considered and what a human reviewer changed or confirmed; and confirmation that the vendor's process holds up under the adjuster-licensing rules covered in Part 2, in the states where the claim is being handled. Calling itself "AI-native" doesn't guarantee any of that — the licensing and sole-basis exposure described in Parts 1 through 3 attaches to the carrier and its adjusters regardless of which vendor's tool produced the recommendation.
A closing thought: human review isn't always the safer check
Every law and rule described across Parts 1 through 3 of this series rests on one assumption: a human reviewer reduces error relative to an algorithm acting alone. For the hard majority of claims work — coverage disputes, credibility calls, contested facts — that assumption may be right but in truth we are not aware of large-scale testing of that assumption.
But a narrower category of claims decisions rests on inputs that are simply verifiable: a policy limit stated in the contract, a total-loss valuation built from clean comparable data, a coverage question that reduces to arithmetic once the facts are settled. For that category, AI review may be more accurate and consistent. People get tired, apply a different standard to the fortieth file of the day than the fourth, and miss a detail buried in a long record; a well-built algorithm applies the same verified inputs the same way every time, with no drift across cases.
The laws described in this series don't distinguish between these two kinds of decisions at all — they all reach for the same fix, a human in the loop, regardless of whether the decision underneath is a judgment call or a math problem. That's not a flaw in what's been passed so far; the law is still catching up to a distinction that only starts to matter once the technology is good enough to expose it. But as adoption grows, regulators will eventually have to answer a question the current generation of statutes doesn't ask: is having a human in the loop genuinely better for consumers or does it add inconsistency and time lags?
Sources: Utah S.B. 319 (2026 General Session), Health Insurance Preauthorization Amendments, enrolled copy (le.utah.gov); Florida CS/CS/HB 527 (2026), bill text, bill history, and House staff analysis (flsenate.gov).
This is Part 4 of 4, the close of this series on AI and claim denials.
