The agentic claims desk: an operating model

1. Where the time goes

Ask an adjuster to narrate a day and most of it is assembly. Reading the policy as it stood on the loss date — not as it stands today — and the endorsements that applied. Finding the documents that support the damage figure and noticing the ones that don’t exist yet. Working out what the reserve should rest on, and justifying it to someone with more authority. Knowing which state’s acknowledgement clock is running and which letter is due. Writing the same status paragraph for the fortieth time.

None of that is the judgment a carrier pays an adjuster for. The judgment is the coverage call, the settlement number, the decision to refer to SIU, the moment to close. The assembly is mechanical in the sense that matters: given the record, a careful reader would reach the same facts every time.

2. The division of labour

An agentic desk splits the work along that line.

• Agents assemble and prepare. One agent per job — intake, coverage, reserving, fraud, subrogation, litigation, medical review, correspondence, the supervisor’s brief, closure — each reading the record when the claim changes and preparing its piece: a coverage position with the forms cited, a reserve with its basis, a letter, a referral, an approval brief.
• People decide. What an agent prepares waits for the person who holds the authority, who accepts it, changes it, rejects it or asks a question.
• The record learns. Every answer is written to the claim’s decision record and counted. Over time the counts say which agent, on which line, for which action, the desk agrees with — and that is the only basis on which an agent should be allowed to do more.

The adjuster’s work changes from building each claim to reviewing it. That is a bigger change in management than in technology, and section 6 deals with it.

3. Authority is earned, per action

“How autonomous is the AI?” is the wrong question; the right one is how much a specific agent may do with a specific action on a specific line. A useful scale has four levels:

Three things cap the level. The carrier sets a ceiling for the whole desk; ask-first is the sensible place to start. Some actions carry a hard cap whatever the evidence says, because a person must own them: closing a claim, filing an SIU referral, deciding an approval. And each agent’s own level per line and action moves only on agreement — promoted when a manager accepts a recommendation backed by a high agreement rate over enough answers, demoted at once when agreement falls.

4. The controls that make it defensible

A regulator, an auditor or the carrier’s own model-risk function will ask the same questions. Each needs a mechanism, not a policy statement.

• Where did that number come from? Every figure an agent states should be traceable to a fact in the record. The simplest enforceable rule is blunt: a model answer that contains a number not present in the facts it was given is discarded, and the rule-based finding stands, labelled as such.
• Who decided? The decision record keeps every run, rule hit, agent finding and human answer on the claim, with the actor, the level it acted at and on whose behalf.
• What did the model see and say? A trace per agent step: the prompt version, the model, tokens, cost, time and output.
• What about personal data? Mask it before anything reaches a prompt, and run the model in the carrier’s own account.
• What stops it going wrong at scale? Per-agent budgets, a circuit breaker on each model, a kill switch per agent, and the ceiling.
• Can it make things worse? Constrain direction where it matters: a coverage position the agent proposes can move toward caution, not away from it.

5. What to measure

Measure the desk, not the model. Benchmarks of model accuracy say little about whether claims are handled better. Useful measures are the ones a claims leader already owns, tracked before and after:

• Agreement — the share of prepared decisions accepted unchanged, per agent, line and action. It is the measure that governs authority.
• Time to the coverage position and time to the first reserve from notice.
• Reserve development — how far opening reserves move before close, and how often they move in small steps.
• Recoveries opened against those available, and statutes missed.
• State-clock breaches per hundred claims.
• Adjuster capacity — open claims per adjuster at constant quality.
• Model cost per claim — recorded per step, so it can be stated rather than estimated.

6. Adopting it

1. Shadow. Agents run at draft on one line. Adjusters work as before; the desk reads what the agents would have prepared and where it differs.
2. Ask first on one line. The agents’ preparations become the adjusters’ starting point. Agreement starts to accumulate.
3. Widen. Add lines one at a time, each with its own line handling, until the evidence is there across the book.
4. Raise specific actions. Where agreement is sustained, a manager accepts act-and-tell for named, reversible actions — a routine status letter, a payment inside a small authority — each with a pull-back window.

The management change is in step 2: supervisors stop reviewing files and start reviewing answers, and adjusters are measured on the quality of their decisions rather than the volume of their typing. Plan for that before the software.

7. How InsightCW implements it

InsightCW is built on exactly this model: ten claims agents with the hard caps above, an L1–L4 dial with an L2 ceiling out of the box, prepared decisions answered with Accept, Change, Reject or Ask, a decision record per claim, the grounding rule, traces, budgets and PII masking — connected to InsightUW for the policy at the loss date and to InsightCede for cessions. The docs describe each mechanism; the blog follows claims through it end to end.