Claims triage agents cut handling time from 15 days to 35 minutes

The insurer processed 400,000 claims a year through a first-notice process that averaged fifteen days from intake to handler assignment. Backlogs grew at each renewal peak, customer satisfaction fell, and a regulatory review had flagged inconsistent classification.

Two previous pilots had produced convincing demos but were not allowed into production: neither could explain its decisions, respect the approval thresholds in the claims manual, or produce an audit trail.

We started with a three-week Opportunity Assessment that established the baseline, the decision boundaries, and the governance classification. Triage affecting individual outcomes was treated as high-risk under the EU AI Act from the outset, so oversight design and logging were requirements, not features.

Delivery ran in two-week increments. The first slice handled one claim type end-to-end in shadow mode; evaluation compared agent decisions with handler decisions on 12,000 historical claims.

A small set of specialised agents — intake extraction, coverage check, fraud signal enrichment, and routing — coordinated by an explicit state machine rather than a free-running loop. Every action the agents may take is an allow-listed, typed tool; monetary and liability thresholds trigger human approval.

The system runs on the client's private OVHcloud environment with a model gateway that routes extraction to a small open-weight model and reasoning to a commercial model, keeping cost per claim below €0.40. Traces, evaluations, and cost are visible per claim type.