Claims triage copilot
RAG over policy documents with adjuster-in-the-loop approvals.
Duration
11 weeks to production
Engagement
Discovery → production AI
Industry
Insurance
Client
Insurer
−38%
Triage time
Median time to first decision
96%
Citation accuracy
On held-out evaluation set
11 wks
Time to production
Workshop to live desks
91%
Adjuster adoption
Weekly active among trained users
Project brief
What this project was really about
An insurer’s adjusters spent hours hunting policy clauses while claim queues grew. We shipped a private RAG assistant with citation-first UX, evaluation harnesses and human approval gates — production-ready in eleven weeks.
01 — Challenge
Adjusters lost time searching unstructured policy documents. Early AI pilots hallucinated clauses and could not survive compliance review.
- Scattered PDFs and legacy claim notes
- No evaluation suite for answer quality
- Risk of uncited model recommendations
- Need for full audit trails on every suggestion
02 — Solution
Private RAG over access-controlled corpora, an LLM gateway with logging, and a UI that refuses to recommend without a citation the adjuster can verify.
- Ingestion pipeline with PII redaction
- Vector store with row-level security
- Eval harness wired into CI
- Human-in-the-loop approval workflow
How we delivered it
A phased programme with evidence at every gate.
- 01
Corpus & risk workshop
Defined refusal policies, citation requirements and which claim types were in scope for v1.
- 02
Retrieval & evals
Built ingestion, retrieval quality baselines and regression suites before expanding tool access.
- 03
Copilot UI
Adjuster workspace with cite-or-quiet behaviour and mandatory approval before claim action.
- 04
Production hardening
Latency budgets, cost controls, audit logs and a phased rollout across claim desks.
04 — Architecture
System shape
- 01Document ingestion and chunking pipeline
- 02Vector store with row-level security
- 03LLM gateway with audit logs
- 04Human approval workflow
- 05Evaluation suite in CI
- 06Observability for cost and latency
Deliverables
What shipped
- Production claims copilot
- Private document corpus with ACL
- Evaluation harness and dashboards
- Adjuster training and playbooks
- Compliance review pack
Stack
Client voice
“Finally an AI tool our compliance team did not ask us to turn off.”
James Okonkwo
Head of Claims Operations · Insurer
Related projects
Have a programme that needs the same discipline?
Tell us about the constraints, the risk, and the outcome you need. We will respond with a clear engagement shape — not a slide deck of buzzwords.