AI Agents for Insurance Claims: First Notice of Loss to Adjuster

The safest place to put AI agents in insurance claims is the front door. First notice of loss, document checks, coverage lookups, and routing to the right adjuster.

The adjuster still decides coverage, liability, and payment. The agent prepares the file so the adjuster starts with a complete one.

Here is where claims time goes, what an agent can do at intake, what must stay human, and a narrow pilot you can run.

Where claims time goes

Before an adjuster makes a judgment call, a lot of work happens that is not judgment. Someone reads the first notice of loss (FNOL). Someone checks the policy was active on the date of loss. Someone finds the missing police report, photo, or estimate. Someone decides which queue the claim belongs in. Someone writes back to the customer asking for the thing they forgot.

That is intake work. It is repetitive, rule-heavy, and spread across email, portals, phone notes, and PDFs.

McKinsey’s “Claims 2030” article, published in May 2022, draws a line I agree with. For simple claims with predictable patterns, the technology for full straight-through processing already exists. For the most complex claims, humans stay in charge because those need empathy, expert judgment, and more data than a model has. The article describes AI that segments and routes each claim at first notice and drafts an initial damage estimate from photos.

That article is a consultant’s view of the future, and it is four years old as of this writing. I use it for the shape of the split, not for any number.

My read: the pain is in the middle of the file, between “claim opened” and “adjuster can act.” That is where an agent helps first.

Intake and document checks

An intake agent reads what came in and builds a clean claim record. It works from the customer’s message, form, photos, and attachments.

What it can do:

• Pull structured fields from free text: date and place of loss, parties, vehicle or property, what happened, injuries mentioned.
• Check completeness against a checklist for that claim type. Missing photos, no police report number, no repair estimate.
• Ask the claimant for what is missing, in plain language, using approved templates.
• Flag inconsistencies: date of loss in the form differs from the date in the email, or the photo metadata does not match.
• Summarize the file for the adjuster in a fixed format, with links back to each source.

Documents are the hard part. Scanned forms, phone photos of estimates, and faxed reports break simple pipelines. I cover the trade-offs in document extraction: LLM vs OCR. For claims, my default is to extract with a model, then validate each field against a rule or a second source before it goes into the claim record.

Two cautions:

1. Photos can be altered. Editing tools make it easy to add damage to a real photo. Intake should treat images as claimant-supplied evidence, not truth. Keep the original file and flag odd cases for human review. I do not have a trustworthy statistic on how common this is, so I will not quote one.
2. An agent that fills gaps by guessing is worse than none. If the date is missing, the agent asks. It does not infer.

Coverage lookups

Once the claim record exists, the agent can look up the policy and answer simple factual questions:

• Was the policy in force on the date of loss?
• Is the insured peril listed?
• What are the limits and deductible?
• Are there endorsements or exclusions that touch this loss type?
• Is there a prior claim on the same loss?

This is retrieval, not decision-making. The agent reports what the policy documents say and shows the clause. It does not say “covered” or “not covered.”

That distinction matters. In February 2024, a Canadian tribunal held Air Canada responsible for wrong policy advice from its website chatbot and rejected the argument that the bot was a separate entity, according to a McCarthy Tétrault summary. It was a small claims ruling and is not binding precedent. But the principle is simple: if your system tells a customer something about their coverage, you own that statement.

So I recommend two rules for any coverage lookup:

• The agent quotes the policy text and the page or clause, every time.
• Anything stated to the claimant about coverage is either pulled word for word from approved language or reviewed by a person first.

Triage and routing

Triage is where an agent saves the most adjuster time. It reads the structured claim and assigns:

• Line and type: auto glass, auto collision, water damage, theft, liability.
• Complexity band: simple, standard, complex, based on rules you write. Injury, litigation, large loss, or a coverage question moves a claim up.
• Queue and handler: by skill, license, region, and workload.
• Priority flags: vulnerable customer, catastrophe event, statutory deadline coming.
• Fraud-review flag: only as a flag for the special investigations team, never as a decision.

Write the routing rules down as a table. The agent follows the table, and a person owns changes to it.

Signal in the file	Route to
Injury mentioned	Complex queue, human review of the whole file
Attorney named	Complex queue, notify supervisor
Loss amount over your authority threshold	Senior adjuster
Coverage question found in lookup	Coverage specialist
Missing required documents	Back to claimant, hold for follow-up
Simple, complete, within a clear band	Standard queue with the summary attached

Misrouting has a cost, but a fixable one. Make the agent log its reason for each route so adjusters can correct it and you can measure the miss rate.

What stays with the adjuster

I would keep these decisions human for any first deployment:

• Coverage decisions. Covered, partly covered, or denied.
• Liability and fault.
• Reserve setting and payment amount.
• Any denial or reduction.
• Settlement negotiation.
• Fraud determinations.
• Contact with injured or distressed claimants beyond routine status updates.

This is also where regulators are heading. The NAIC says AI in insurance is more likely to support human workers than replace them, and insurers stay responsible for compliance, fairness, and accuracy. At the state level, Florida’s HB 527 in 2026 would have barred AI as the sole basis for denying or reducing a claim and required that those decisions be made by qualified human professionals. According to the Florida Senate record, the House passed it 108-0 on March 5, 2026, but it died in the Senate Rules Committee on March 13. It is not law. It does show the direction of legislative interest, and I would design as if a version of it will return.

Agents can also do this work badly if they run unchecked. Anthropic’s guidance on building agents suggests pausing for human feedback at checkpoints and setting stopping conditions. For claims, the checkpoint is simple: every step that affects money or coverage goes to a person.

Audit trail and fairness

The question that matters is not whether you used AI. It is whether you can show how a claim was handled and whether it was handled fairly.

The NAIC adopted its Model Bulletin on AI use by insurers in December 2023. As of March 2025, 24 jurisdictions had adopted it, according to a Quarles tally. Some states run their own frameworks instead. Check your own state department, because requirements differ. Quarles summarizes the main expectations:

• a written program for responsible AI use, with governance and internal audit
• clear accountability across business functions
• controls matched to consumer risk, including data quality, bias, testing, and record retention
• vendor oversight, with contract terms for audit rights and cooperation with regulators
• readiness to produce documentation in an exam or market conduct review

I am not a lawyer, and this is not legal advice. Have compliance read the bulletin and your state’s version.

In practice, an audit trail for intake should record, per claim:

1. the inputs received, with timestamps, kept unaltered
2. each field the agent extracted, the source it came from, and any validation result
3. each policy lookup, with the clause returned
4. the triage decision, the rule that fired, and the confidence or flags
5. every message sent to the claimant, and the template used
6. every human override, with the person and the reason
7. the model, prompt, and rule versions in effect

I wrote a full list in what to log for an AI agent audit trail.

For fairness, test the intake agent’s outputs by segment. Does it route claims from some regions, languages, or channels to slower queues? Does it ask some claimants for more documents than others? Run those checks before launch and monthly after.

A narrow first pilot

Here is a pilot I would be comfortable with. The example is hypothetical.

Say a regional carrier has a 25-person auto claims team. Pick one line, such as auto glass or minor collision with no injury. Run the agent in shadow mode first.

Weeks 1 to 2: shadow. The agent processes each new claim in the background. It extracts fields, checks completeness, looks up coverage, and proposes a route. Nothing it produces reaches a claimant or changes a queue. Adjusters work as usual. You compare its output to what the humans did.

Weeks 3 to 4: assist. The agent’s summary and route go to a human intake clerk, who approves or corrects. The agent may draft requests for missing documents, and a clerk sends them.

Decide. Only if the numbers hold do you let the agent send standard missing-document requests directly, from approved templates.

Pick one hero metric before you start. Good candidates:

• time from FNOL to adjuster-ready file
• share of files complete at first handoff
• routing accuracy, measured against a senior adjuster’s review of a sample
• adjuster minutes of intake work per claim

Set stop rules in advance. For example, stop and review if routing accuracy falls below your bar, if any claimant message goes out that was not from an approved template, or if a coverage statement lacks a clause citation.

To estimate value, use the framework in AI agents for back-office ROI. Count adjuster time saved, but also count rework avoided and days to first contact. For the wider picture on regulated finance workflows, see AI agents in financial services.

Bottom Line

Start at the front door. Use an agent for intake, document checks, coverage lookups, and triage. Keep coverage, liability, payment, and denials with the adjuster.

Log every field, lookup, route, message, and override. Test for fairness by segment.

Next step: pick one simple line, run the agent in shadow mode for two weeks, and measure it against your adjusters.

FAQ

How is AI used in claims? Common uses are intake, document reading, coverage lookup, triage and routing, status updates, and fraud flagging. These are the repetitive, rule-heavy steps. Judgment calls like coverage and payment mostly stay with adjusters.

Can AI process claims end to end? For simple claims with predictable patterns, McKinsey says the technology for full straight-through processing exists. Whether you should depends on your regulators, your risk appetite, and your data. I recommend starting with intake and triage, with a human on every decision that moves money.

What must an adjuster still decide? Coverage, liability, reserves, payment amounts, any denial or reduction, settlement, and fraud findings. An agent can prepare the file and cite the policy. A person should own the decision, especially for denials.

What audit trail do regulators want? Regulators expect documentation of how AI is governed and controlled, and that you can produce it in an exam. Per claim, keep the inputs, extracted fields, policy clauses used, routing rule, messages sent, human overrides, and model and prompt versions. Check your state’s version of the NAIC bulletin.

What is a safe first pilot? One simple line with no injuries. Run the agent in shadow mode for about two weeks, then in assist mode with a clerk approving each output. Track one metric, such as time from first notice to an adjuster-ready file.

Will an agent catch fraud? It can flag patterns for your special investigations team, such as mismatched dates or reused photos. It should never decide fraud. Treat a flag as a reason to look, not a finding.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect