Triage - AI Claims Copilot

Car-insurance claims run on two people's judgement in sequence. An agent looks at the damage photos, prices the repair by eye and experience, and a senior adjuster re-reviews it before anything is authorised. It's slow, it's inconsistent (two agents will price the same dent differently), and it spends senior time on a scratched bumper in exactly the same way it does on a write-off.

Vision models can already identify damage in a photo. What is less clear is how much of the review and approval process can safely run without a person, and how the system should recognise the claims that still need one.

I built Triage, a working prototype of an adjuster's copilot. It reads the damage photos, drafts an itemised estimate, and scores its own confidence on each line. Then it routes the claim into one of four lanes: auto-approve, assisted review, escalate, or total-loss review.

It is designed to support adjusters, not to replace them. Confidence is scored for each line item separately, because averaging the scores into one number would hide the panel that most needs a person to check it. A fraud signal overrides every other rule, so a claim that looks straightforward but has a suspicious photo still goes to a person. The thresholds that decide which lane a claim goes into are settings on the page. These thresholds are effectively the insurer's risk policy, so the insurer should set them deliberately rather than have them fixed in the code.

Alongside the prototype I worked out the eval plan and the feedback loop behind it. The most important part is easy to overlook. Auto-approved claims produce no feedback on their own, so a wrong estimate that nobody complains about looks the same as a correct one. The plan catches these with a regular audit of a sample of auto-approved claims, and only the results of that audit can justify widening what gets auto-approved.

I prototyped the whole thing hands-on in Claude Code. The vision step is simulated with a clickable diagram, because the product decisions I wanted to show all come after the damage has been identified.

A working demonstration of where automation should stop in a regulated, high-liability process, and how the insurer can audit that limit as well as adjust it. In the live tool you can move the thresholds and see claims move between lanes.