AI Claims Adjusting 2026: Adjuster Licensing and the Unfair Claims Practices Acts

Two Duties Automation Struggles With

The first is investigation. Unfair claims practices acts prohibit refusing to pay without a reasonable investigation based on all available information. Automated pipelines are excellent at processing what was ingested and structurally blind to what was not — the adjuster note, the uploaded PDF nobody parsed, the second photograph set sent by email, the phone call in which the claimant explained the sequence of events. When a file is closed on a model output, the question at examination is not whether the output was accurate. It is what information existed in the file and whether it was considered.

The second is explanation. When a claim is denied, the claimant is entitled to a statement of the basis: the policy provision, condition or exclusion relied on. This is where a scoring architecture and a statutory duty collide most directly. A model produces a probability; the statute demands a contractual reason. Someone must perform the translation, and if that translation is itself templated by claim type, the letter describes a category rather than this claim — which is precisely the complaint a market-conduct examiner is reading when they pull a sample of denials and find the same paragraph.

Seven Claims Surfaces, Graded

First notice of loss intake and document classification

Low risk, high value

Capturing and organising the file does not decide anything. Keep the claimant's own description verbatim — it is the record of what was reported and when.

Coverage lookup and policy-provision retrieval

Support, if it cites

Surfacing the applicable provision helps the adjuster and improves denial letters. It must link to the actual policy language rather than a paraphrase of it.

Photo-based damage estimation

Valuation, with a known failure mode

Image estimates systematically miss what is not visible — substrate damage, hidden water, pre-existing condition. The estimate is a starting figure, and treating it as final is where supplement disputes come from.

Fraud scoring and SIU referral

Not a denial, but it delays like one

A referral that pauses a claim without triggering any notice recreates the effect of a denial with none of the process. Bound the pause and measure how long flagged claims actually take.

Injury severity or treatment-duration scoring

High risk

Population-derived predictions about an individual's recovery are the classic bad-faith exhibit, and the discovery question — what were you optimising — has no comfortable answer.

Automated denial or partial denial

Needs a named human and a policy provision

The letter must state the specific basis relied on. A score is not a basis, and a reviewer who cannot practically reach a different conclusion is not a decision-maker.

Settlement offer generation and negotiation copy

Watch the words, not the number

Unfair-practices statutes reach misrepresenting policy provisions and compelling litigation by offering substantially less than eventually recovered. Generated persuasive text is where a misstatement of coverage appears.

The Asymmetry Nobody Configures For

Claims automation is usually tuned as a single accuracy problem with one threshold. But the two errors it can make are not comparable. A wrong automatic approval costs the insurer the amount of the claim, is discovered internally, and is correctable next quarter. A wrong automatic denial costs a claimant the remedy they paid premiums for, is discovered only if that person has the resources to push back, and converts into regulatory and bad-faith exposure that is an order of magnitude larger than the claim.

The design implication is that the two paths should not share a confidence threshold, a review rate or an audit sample. Automatic approval can run at a low bar because the downside is bounded and internal. Anything trending toward denial should carry a high bar, a named reviewer with real authority to disagree, and a sampled quality review that reads the file rather than the output. Most pipelines are built the other way round, because approvals are the volume and denials are the exception — so the exception path gets the least engineering.

What Gets Requested in an Examination

The insurance-AI bulletin adopted across many states is a governance document, and reading it as a document-request list is the fastest way to see whether a programme is ready. Expect to be asked for the AI systems inventory and what decisions each model touches; the written programme and who owns it; risk assessments proportional to consumer impact; validation and drift monitoring, including testing for unfair discrimination; third-party model controls and the contract terms supporting them; and evidence that consumers affected by an AI-influenced adverse decision had a route to a human.

Every one of those is an artefact, not an opinion. The typical shortfall is not that an insurer disagrees with the standard — it is that the work was genuinely done in a series of meetings and nothing was written down, so there is nothing to produce. Writing the inventory is a week of work that converts an unanswerable request into a routine one.

A Defensible Configuration

No adverse determination without a named human and a cited provision. The reviewer needs the file, the time and the authority to disagree, and the letter needs a real contractual basis rather than a rendered category.
Split thresholds by direction. Approvals and denials are different risk profiles and should never share a confidence bar.
Log what the model did not see. Unparsed attachments, unprocessed notes and unread correspondence are the evidence of an incomplete investigation. Make the gap visible in the file rather than invisible.
Version and date every threshold change, with a rationale. This is the exhibit either way; the only question is whether it reads as engineering or as a payout decision.
Measure cycle time for flagged claims. A fraud referral that quietly triples the time to payment is a prompt-pay problem wearing an investigation label.
Negotiate model transparency at procurement. The duty to explain runs to the insurer regardless of what the vendor will disclose, so the disclosure right belongs in the contract.

Frequently Asked Questions

Does an AI system need an adjuster licence?

Licences are issued to people, so the literal answer is that software cannot hold one — which is why the useful question is whether the activity being performed is adjusting, and if so, which licensed human is performing it. Most states license adjusters and define the activity functionally: investigating, negotiating or settling claims on behalf of an insurer or an insured, for compensation. Some states exempt salaried employees of an insurer, some license company adjusters as well as independents and public adjusters, and the exemptions vary enough that a multistate programme cannot be designed from one state's rule. What is consistent is that automating the activity does not remove it from the definition. If a system investigates, values and settles, an adjusting activity happened, and a regulator will ask who was licensed to do it and what they actually did.

What does 'reasonable investigation' mean when the investigation is a model?

Unfair claims settlement practices acts, which nearly every state has adopted in some form, prohibit refusing to pay a claim without conducting a reasonable investigation based upon all available information. That standard was written with a human file-handler in mind, and it has three components an automated pipeline tends to fail. Completeness: the model sees the fields ingested, not the documents in the file or the context in a phone call. Responsiveness to submitted evidence: a claimant who uploads a contractor's report expects it to change the outcome, and a pipeline that scores photographs will not have read it. Individual consideration: a model output derived from population-level patterns is, by construction, an inference about similar claims rather than an investigation of this one. None of these are fatal, but each has to be answered by design rather than assumed.

Can a claim be denied automatically?

This is the sharpest line in the category, and jurisdictions are moving in the same direction on it: an adverse determination should be made or adopted by a qualified person who can explain it. Several regulators have said so in the health-utilisation-review context explicitly, and the reasoning transfers to property and casualty. Two operational realities matter more than the formal rule. Denial letters must state the specific policy provision, condition or exclusion relied on, and a model score is not a provision — someone has to map the output to a real contractual basis. And the volume asymmetry is dangerous: an automated approval that is wrong costs the insurer money, while an automated denial that is wrong costs the claimant their remedy and the insurer its bad-faith exposure. Those two errors are not symmetric and should not share a threshold.

How does automation create bad-faith exposure specifically?

Bad-faith claims turn on the reasonableness of the insurer's conduct and the insurer's knowledge, and automation generates unusually good discovery on both. A model has a configuration history, a threshold that someone chose, a version that changed on a date, and performance metrics that someone reviewed. A plaintiff's lawyer will ask what the approval rate was before and after the threshold moved, whether anyone modelled the effect on claimants, and what the internal name of the metric being optimised was. A file in which a human adjuster reached a defensible conclusion is a difficult case; a file in which a threshold change reduced payouts by a measurable percentage in the same quarter it was deployed is a different kind of case, because it looks like a decision about a population rather than about a claim.

What did the NAIC model bulletin actually ask insurers to do?

The insurance-AI bulletin adopted by a large number of states is a governance instrument rather than a prohibition, and it is worth reading as a list of things a regulator may request. It expects a written AI systems programme owned by senior management, an inventory of where AI is used and for what decisions, risk assessment proportional to consumer impact, validation and ongoing performance monitoring including for unfair discrimination, controls around third-party and vendor models, documentation retained and producible, and a mechanism for consumers affected by an adverse AI-influenced decision. Almost every element is documentation of decisions a well-run programme already made. The common failure is not disagreement with the standard but absence of artefacts: the work happened and no one wrote it down, so nothing can be produced when the market-conduct examination asks.

Are vendor models the insurer's problem?

Yes, and that is the most frequently mispriced risk in insurtech procurement. The regulated entity is the insurer, the duties run to the insurer, and outsourcing the model does not outsource the obligation. The recurring gap is contractual: the insurer needs the right to obtain validation results, to be told when the model changes materially, to have data produced for a regulator or in litigation, and to know what data the vendor retains and reuses. Vendors frequently decline several of these on trade-secret grounds, which is a commercial negotiation with a regulatory consequence — an insurer that cannot explain a decision because its vendor will not explain the model still owes the claimant an explanation. The correct time to discover this is procurement, not the first market-conduct exam.

Which claims-handling surfaces are the lowest risk to automate?

The ones that accelerate the file without deciding it. Intake and first notice of loss capture, document classification and routing, duplicate detection, coverage-page retrieval, status communications, and reserve-setting support are all improvements that leave the determination with an adjuster. Fraud referral is a middle case: flagging a file for investigation is not a denial, but a flag that silently slows a claim produces the same effect as a denial while generating none of the required notices, so the delay it causes should be measured and bounded. The high-risk surfaces are denial, valuation of a total loss, injury-severity or duration scoring, and any threshold that routes claims to a lower-touch path based on a predicted characteristic of the claimant rather than a fact about the loss.

The Denial Letter Sample

Pull fifty denial letters issued on files the model touched. For each, find the specific policy provision cited and check whether the facts recited are facts about that claim or boilerplate for that claim type.

Then find the human who decided. If the same name appears on all fifty and the average time between assignment and decision is under a minute, you have documented that the review was a signature — and you have documented it before a plaintiff's lawyer did.

The Threshold You Moved Last Quarter Is the First Exhibit