Case 02 / Insurance
An AI triage desk for an insurance operations team
Claims arrive as messy documents. The product reads them, ranks them, and puts a human in front of the ones that matter.
- Engagement
- AI product build
- Status
- Live in operations
- Duration
- Two release cycles
- Team
- Cyrque, end to end
01 / The challenge
A queue that grew faster than the team reading it.
Every claim came in as scans, forms and free text, and a small operations team read all of it in arrival order. The work was not hard, it was undifferentiated: the urgent and the routine looked identical until someone opened them.
02 / Our role
We designed the model into the workflow instead of beside it.
We set the accuracy and cost targets first, then built extraction and ranking into the tool the team already worked in, with a review path for anything the model was unsure about.
Document extraction
Structured fields pulled from scans and forms, with confidence carried through to the interface.
Ranking and routing
A queue ordered by urgency and value, with the reasoning visible to the person acting on it.
Human review
Low-confidence cases escalate rather than resolve silently, and corrections feed back into evaluation.
Evaluation harness
A test set and metrics the team owns, so a model change can be judged before it goes live.
03 / The outcome
The team reads less and decides more.
Triage moved from arrival order to priority order, and the operations team spends its attention on the exceptions instead of the sorting.
04 / Impact
The numbers we track with the client.
Reduction in time from claim arrival to a human decision.
Claims read and structured without manual entry.
Measured against the reviewed set the team owns.
Cases cleared per person per day versus the old queue.
Figures are placeholders pending client sign-off.
05 / Start here