An AI Mission for Insurance: Claims Adjudication

An insurer sells a promise, and adjudication is the moment the promise is tested. The part of that moment a machine can carry is the fact-finding underneath it — never the determination itself, which has to remain a decision a person made and can defend.

A claim file arrives on an adjuster's desk as a small archive of other people's worst week. There is a first notice of loss taken over the phone by someone who was upset and not especially organised, a set of photographs shot at odd angles, an invoice from a contractor or a provider written in a shorthand only that trade uses, a policy document running to dozens of pages with endorsements that modify clauses buried elsewhere in it, and a claims history in a system that does not speak to the document store. The adjuster's job, stated plainly, is to work out whether what happened is a thing the policy pays for. Before they can even begin to answer that, they have to spend most of their day answering a much less interesting set of questions: what coverage was actually in force on the date of loss, what the insured reported and when, what the documentation in front of them does and does not establish, and which of the policy's conditions and exclusions are even in play. Only after all of that is settled does the question that requires a human show up — and by then the day is largely gone.

That distribution of effort is the thing worth staring at, because it explains why claims adjudication has resisted every generation of automation aimed at it. The determination is a small act of judgment sitting on top of a very large act of retrieval and reconciliation, and the industry has repeatedly tried to automate the judgment while leaving the retrieval to people. It should have been the other way around. The facts a coverage determination turns on are, almost entirely, the kind of thing that can be found, extracted, cross-checked, and assembled by software that reads. The determination itself — the reading of policy language against those facts, and the decision that follows — is the part that has to stay with a person, because it is an act that will have to be explained later, to a claimant, to a supervisor, to a regulator, and sometimes to a court.

The fact-finding is not the decision, and the two have very different requirements

It helps to separate an adjudication into its parts and notice how differently they behave. Establishing what coverage was in force is a lookup problem across a policy administration system, an endorsement history, and a payment record — tedious, error-prone when done by hand at volume, and entirely mechanical in nature. Establishing what was reported is a reading problem: the first notice of loss, the supplemental statements, the correspondence trail, all of which have to be reconciled into a single coherent account with its inconsistencies flagged rather than smoothed over. Establishing what the documentation supports is a comparison problem: does the estimate describe the same damage the photographs show, does the treatment record correspond to the dates claimed, is the vendor invoice internally consistent, and where does the file simply go quiet on something the policy makes relevant. None of these three activities requires anyone to interpret policy language. All three of them consume the overwhelming majority of the time an adjuster spends on a file.

The determination behaves nothing like that. It requires reading a contract's words in light of how the insurer has read those same words in comparable situations, weighing what the record establishes against what it merely suggests, and accepting responsibility for a conclusion that may be adverse to the person who bought the policy. That last part is not a technicality. A denial, or a partial payment, is an adverse action taken against someone who is frequently at the lowest point they have been in years — after a fire, a crash, a diagnosis, a death in the family. The person on the other end is entitled to know why, in terms specific to their file, and the insurer is obliged to be able to say why in a way that holds up under scrutiny. That obligation is not satisfied by a system that produced an output; it is satisfied by a named human being who examined a record and reached a conclusion they can articulate and stand behind.

So the design rule follows directly from the nature of the work, and it should be stated without hedging: the AI never denies a claim, never approves a claim, and never adjudicates one. It does not score coverage, it does not recommend a disposition dressed up as a "suggested outcome" that an adjuster is expected to rubber-stamp, and it does not carry any authority over the money. What it does is establish the factual record the determination will rest on, present that record with its gaps and contradictions visible, and hand it to the person whose decision it is. Anything beyond that line is not efficiency; it is the quiet transfer of an accountable act to something that cannot be held accountable.

Defensibility is a property of the record, not of the outcome

This is a different problem from the one that dominates most conversations about AI in claims, which is usually about throughput — intake, triage, routing, and the handling of exceptions that fall outside a straight-through path. That work matters, and it is genuinely a coordination problem. Adjudication is not. The question here is not how fast a file moves but whether the decision at the end of it can be defended six months later, when nobody remembers the file and all that survives is what was written down.

Defensibility, understood properly, is not about being right. It is about being able to show the basis on which you decided, in the order you decided it, with the sources attached. An adjudication is defensible when someone can reconstruct which policy version applied and why, what the claimant reported and when, which documents were considered, what each of them established, where the record was incomplete and what was done about it, and how the person deciding moved from that record to their conclusion. In most operations today that reconstruction is difficult, not because anyone acted badly but because the reasoning lived in an adjuster's head and the record lives scattered across four systems and an email thread. The file shows the outcome and hides the path.

An AI mission built for adjudication is, in this framing, primarily a machine for producing that path. Specialist agents work the retrieval and reconciliation across the policy administration system, the document store, the claims history, and the correspondence record, each returning not just an answer but the source it came from and the confidence it has in it. A reasoning layer assembles the results into a coverage-relevant factual summary — what is established, what is asserted but unsupported, what is missing, what conflicts — and explicitly declines to draw the conclusion. Human-in-the-loop here is not a checkbox appended to an otherwise automated pipeline; it is the point of the pipeline. The system's entire output is an input to a person's decision, and the audit trail it generates as a by-product is the artefact that makes the decision explainable long after the fact.

The workforce shift is real, but it stops at the decision

Insurers evaluating this should be sceptical, and the market gives them reason to be. Gartner has predicted that more than forty percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls, alongside a good deal of "agent washing" — familiar rule engines and chatbots relabelled as autonomous. In a regulated adverse-action context that last failure mode is more than a procurement embarrassment. A system that quietly accretes decision authority it was never meant to have, because a suggested disposition became a default and the default became the practice, is a governance failure that shows up years later in a pattern nobody chose.

The more durable version of the idea is narrower and, precisely because it is narrower, far more defensible. Platforms in this space, StudioX's insurance missions among them, are built around the premise that autonomous AI workers should absorb the establishing work and stop cleanly at the boundary of the determination — the same logic that runs through the broader shift toward the autonomous enterprise, where software takes over the connective labour between systems and people retain the acts that carry consequence. The adjuster's day changes enormously under that arrangement: the hours spent hunting for the endorsement history and reconciling the loss description against the estimate largely disappear, and what remains is the reading of the policy against a record that has already been built. What does not change is who decides, and who signs.

The mental model worth carrying out of this is that adjudication has two products, not one. There is the determination, which the claimant experiences, and there is the record of how it was reached, which everyone else — the supervisor, the examiner, the regulator, the claimant's lawyer — eventually experiences instead. Insurers have historically optimised the first and treated the second as paperwork. The opportunity in front of them is to build the second properly and automatically, because a determination is only as good as the record that can explain it, and an insurer that can always explain itself is holding something more valuable than speed.