Insurance adjusters don't guess. They read the claim, pull the policy, check history, apply rules, and decide — with paperwork that shows why. Our new POC automates that workflow, not that judgment call.

Repo: xingai-enterprise-ai-pocs/pocs/claims-multiagent-rag-poc/

The problem with one big prompt

A single LLM call that "reads the claim AND checks the policy AND checks fraud AND decides" is unreliable:

• The prompt becomes huge and untestable.
• You can't audit which step failed.
• The model may deny a claim without quoting an exclusion clause.

We split the job into specialized agents with a supervisor (LangGraph) that owns explicit handoffs.

Claim → Intake → Retrieval (RAG) → Fraud-Check → Adjudication → Audit
              ↘ low confidence → Human Review

What RAG means here

Retrieval-Augmented Generation: before any decision text is generated, we retrieve real document chunks from three separate Chroma collections:

Collection	Contents
policy_documents	Coverage, exclusions, deductibles
claim_history	Past claims per synthetic policyholder
regulations	State handling rules

Every chunk carries document_id, chunk_id, and similarity score. Downstream agents never see naked text without a citation.

Rules before models

Fraud-Check runs deterministic rules first (claim frequency, amount vs limit, new-policy window), then a lightweight narrative pass. High fraud score → escalate to human — we never auto-deny on fraud alone.

Adjudication must cite at least one policy excerpt for APPROVE or DENY. Thresholds live in config/claims_policy.yml — no magic numbers in Python:

• human_review_threshold_usd: 5000
• escalate_risk_score: 0.70

Audit trail

Every agent step appends to SQLite (audit_trail). PII patterns are redacted before logging. One query answers: "Why was claim #123 denied?"

Demo paths (live script)

Three synthetic claims walk stakeholders through different outcomes:

1. APPROVE — $450 windshield on POL-1001, low fraud score, glass coverage cited.
2. DENY — flood damage on POL-2002; denial quotes the flood exclusion clause.
3. ESCALATE — third glass claim in 30 days, or amount over $5,000.

See demo_script.md.

Golden-set eval

Ten synthetic claims with expected actions; CI target ≥ 80% exact-match on decision action. Run:

pytest tests/eval/test_golden_claims.py -v -m eval

Enterprise mapping

This POC validates Enterprise POC ADR-001: supervisor orchestration, append-only audit, human-in-the-loop thresholds, RAG citation discipline.

Production path (not built): Pinecone/Weaviate, PDF intake, field-level encryption, real claims system integration via MCP read tools first.

Positioning: Phase 1 validation of multi-agent RAG + governance — same architecture a production insurance copilot would use, with synthetic data only.