Claims triage that readsthe file first

Client
Meridian Mutual
Sector
Insurance
Delivered
Span 01 — Meridian Mutual

In one paragraph

A triage engine that reads an incoming claim, its attachments and its history, then routes it with a recommended path and a reason an adjuster can argue with.

Measured afterwards

91%
Agreement with senior adjusters
9d → 4.5d
Median cycle time
7,400
Claims triaged per month
68%
Automated without review

The problem

Every claim arrived as a PDF bundle and sat in one queue. Senior adjusters spent the first hour of each day sorting rather than deciding, and complex claims waited behind simple ones. The median cycle time had grown to nine days and the backlog was being managed with overtime.

What we did

We instrumented the existing queue for three weeks to find where time actually went, then built an evaluation set of 1,400 historical claims scored by two senior adjusters. The engine extracts structured fields from the bundle, scores complexity and fraud signals separately, and writes a short rationale citing the specific pages it relied on. Anything above the confidence threshold routes automatically; everything else goes to a human with the extraction already done. Adjuster overrides are recorded and fed back into the weekly evaluation run.

What changed

Triage now agrees with senior adjusters on 91% of claims and the median cycle time fell from nine days to four and a half. The overtime budget for backlog clearance was closed. The evaluation harness runs on every model change, so a regression is caught before release rather than in a complaint.

Built with

PythonClaude APIFastAPIPostgrespgvectorPrefectRagas