Insurance Claims Agents

Find the time before you automate anything.

Claims operations are measured on cycle time, leakage and reopen rate. Before writing a prompt, instrument where the days actually go on your own book — not on the industry's. In almost every line of business the answer is unglamorous and consistent: the clock runs while the file sits incomplete, waiting on a police report, a repair estimate, a medical record, a proof of ownership, a signature, or a claimant who does not know what was asked of them.

• Handler touch time is a minority of elapsed time. The determination itself is often minutes of work spread across weeks of waiting. Automating the minutes and leaving the weeks alone is how a pilot produces a beautiful demo and a flat cycle-time chart.
• Leakage is paid dollars that policy did not require — a covered-limit error, a missed subrogation opportunity, a duplicate payment, an unapplied deductible. It is a checking problem, which agents are good at, rather than a judgement problem, which they are not.
• Reopen rate is the honest quality metric. A claim closed fast and reopened in six weeks cost more than one closed correctly the first time, and it is the number that automation most easily makes worse while every dashboard improves.
• Segment before you generalise. A glass claim, a total-loss auto claim, a hospital bill and a business-interruption claim share a workflow diagram and nothing else. The playbook below applies to all of them; the automation ceiling differs by an order of magnitude between them.

The completeness engine is the product.

The highest-value agent in claims reads the file, compares it against what this claim type requires, and chases exactly what is missing — in the claimant's language, through the channel they actually answer, with a specific ask rather than a form letter. It never decides anything. It is also, not coincidentally, the piece with the clearest ground truth: either the document arrived or it did not.

• Encode the requirement list per claim type, as data. What must be in the file before this claim can be determined, keyed by line of business, jurisdiction and loss cause. This belongs in a table you can diff and version, not in a prompt — the requirements are the regulated part and they change by state.
• Extraction is a parsing problem before it is a model problem. The inbound file is scans, faxes, phone-camera photos and PDFs of forms. Pick the parser deliberately and evaluate it on your own documents; the comparison in document parsers compared and the patterns in document parsing for RAG are the starting point.
• Every extracted field carries a page-and-region citation. A date of loss with no pointer to where it came from is an assertion, and it will eventually be an assertion that contradicts the document it supposedly came from. Grounding here is not a quality nicety — it is what makes the file defensible. See grounding.
• Chase in one message, not five. The agent's advantage over the standard process is that it can read the whole file at once and ask for everything missing in a single specific request. Three sequential requests over three weeks is the failure mode the current process already has.
• Detect contradiction, do not resolve it. When the estimate says one vehicle and the police report says another, the agent's job is to surface the conflict with both citations and stop. Reconciling contradictory evidence is adjudication wearing a different hat.

Adjudication is where the agent stops — and a denial has to cite a clause.

Coverage determination is the regulated act. Every US state has an unfair claims settlement practices statute, and the recurring themes across them are the ones that constrain your architecture: investigate reasonably, act within defined timeframes, and give the claimant a written explanation with reference to the policy provisions when you deny or offer less than claimed. A model that produces a fluent denial without a traceable clause has produced a compliance defect, however good the reasoning looks.

• A recommendation is not a decision, and the distinction must be real. If the handler approves the agent's recommendation 98% of the time without opening the file, you have built an automated determination with a human decoration on top, and a regulator will read it that way. Measure override rate and review time; if they collapse, either restore friction or accept that you are automating the decision and take on the obligations that come with it.
• Denials and reductions get a structured artifact, not prose. The clause invoked, quoted, with a document reference; the facts relied on, each with a citation; the amount and how it was computed. Generate the letter from that structure. The pattern is structured refusal, and in this domain it is what turns a disputed decision into a record instead of an argument.
• Coverage questions and quantum questions are different automations. "Is this covered?" is interpretation against a contract. "What is it worth?" is valuation against data. The first has far less headroom than the second, and teams routinely build one system for both and get the worst of each.
• Gate on calibrated uncertainty, per action. The threshold at which the agent may close a straightforward glass claim and the threshold at which it may touch a bodily-injury reserve are not the same number. Build the coverage–risk curve per action type — see uncertainty and calibration — and hand the low-confidence tail to a human.
• Payment is irreversible; treat it like it. Recovering an overpayment from a claimant is expensive, slow and reputationally awful. The payment step is gated in your code against limits, deductible, prior payments and payee verification — never on the model's arithmetic alone.

Fraud scoring is the trap. Decline it deliberately.

Fraud detection is the use case every executive asks about first and the one with the worst risk-adjusted return for an LLM-based agent. It is a discrimination-liability surface, an explainability surface and a regulatory surface at once, and the model contributes least exactly where the stakes are highest.

• An adverse inference about a person needs a defensible basis, and "the model flagged it" is not one. Anything that routes a claimant into a slower, more adversarial track is an action taken against them, and you will be asked to explain it in terms that do not include an embedding.
• Proxy discrimination is the specific failure, and it is easy to build by accident. Regulators have been explicit that outcomes matter regardless of intent — Colorado's insurance AI regime, built on SB21-169, requires quantitative testing for unfairly discriminatory outcomes from external consumer data, and it is the template others have followed. Free-text narrative is dense with proxies for protected characteristics: postcode, language, name, phrasing.
• Where an agent legitimately helps is assembly, not scoring. Pulling the claim history, matching entities across files, flagging that the same repair shop and the same medical provider appear in an unusual cluster — these are retrieval and joins, they are auditable, and a human investigator decides what they mean.
• Keep special-investigation referral human-initiated. The agent may surface evidence into an investigator's queue. It may not open a fraud file. That line is cheap to hold and expensive to have crossed.

The rails: governance you will be asked to produce, in writing.

Insurance is one of the few domains where the AI-specific supervisory expectations are already concrete rather than anticipated. The NAIC adopted its Model Bulletin on the Use of Artificial Intelligence Systems by Insurers in December 2023, and a majority of US jurisdictions have since issued it or something close to it. It is short, and it asks for artifacts you should have anyway.

• A written AI systems programme with named accountability. Senior-management ownership, documented controls, validation and testing, and a record of what was tested and when. Map it onto the roles in accountability and roles so it is a live document rather than a PDF written for an exam.
• Vendor AI is still your responsibility. The bulletin is explicit that using a third party does not transfer accountability, which means your diligence, your contractual audit rights and your own testing of the vendor's output. See third-party model and vendor risk.
• The audit trail must reconstruct a decision, not just log events. For any claim, you should be able to produce: what the agent saw, which documents and clauses it cited, what it recommended, who approved it, and what model version and prompt produced it. Decision receipts and audit trails cover the mechanism; the retention period is set by your state, not by your storage budget.
• Claim files contain some of the most sensitive data you hold. Medical records, financial records, minors, and in some lines criminal allegations. Data handling here is HIPAA-adjacent at minimum and often squarely inside it — the constraints in healthcare agents apply to the medical portion of a bodily-injury file whether or not you think of yourself as a healthcare company.
• Timeframes are statutory, so make them alarms. Acknowledgement, decision and payment clocks vary by state and line. An agent that quietly holds a file pending clarification can miss one, which converts a workflow bug into a regulatory finding. Every automated hold needs an expiry and an escalation.

Measure completeness and reopens, never touchless rate.

"Percentage of claims handled without human touch" is the metric the programme will be asked for and the one that most reliably rewards the agent behaving badly. Touchless rate goes up when the agent closes ambiguous claims, skips the chase, and denies at the margin — all of which show up somewhere else, on somebody else's budget, one or two quarters later.

• Completeness at first contact. Of the documents this claim type required, what fraction were requested in the first outbound message? This is the metric that actually moves cycle time and it is scoreable offline against closed files.
• Days-to-complete-file, separated from days-to-close. Splitting the clock at the point where the file became decidable tells you which half of the process your agent is affecting — and stops an unchanged determination stage from taking credit for an improved intake stage.
• Reopen rate and complaint rate, by whether an agent touched the claim. Tag at the claim level from day one; you cannot reconstruct the tag afterwards, and it is the comparison that decides whether the programme survives its first regulatory inquiry.
• Leakage caught versus leakage introduced. Count both. An agent that catches unapplied deductibles and also occasionally misreads a limit has a net number, and only the positive half tends to get reported.
• Keep an adversarial slice permanently. Claimant-supplied documents are attacker-writable in the ordinary commercial sense: a repair estimate is a PDF someone else produced, and an instruction embedded in it should never reach the agent's control path. Build the slice from real submissions and grow it every time you add a channel.

Related: finance agents for the adjacent regulated-money patterns, legal agents for reading a contract as a source of authority, the regulatory landscape for what else is arriving, and adapting a playbook for the method behind this one.