How does AI detect insurance fraud in claims?

AI fraud detection in P&C insurance claims is a layered pipeline that combines deterministic rules, LLM-narrative analysis, and external database lookups. No single layer is sufficient — fraud signals are diverse enough that overlapping detectors are required.

The three layers

1. Rule engine (deterministic). Six core rules fire in milliseconds against the FNOL + claim_history + policy data: - rapid_post_bind — loss reported < 30 days from policy inception (medium); < 7 days (high); < 24h (critical) - round_loss — incurred is an exact round number ($1k / $5k / $10k / $25k / $50k / $100k) → low signal alone, medium when paired with no contractor estimate - prior_similar — same claimant + same cause of loss within 12 months → medium; 2+ priors → high - no_police_report — cause is theft / auto / vandalism but no police_report_number on the FNOL → medium - high_value_low_documentation — incurred > $50k with zero supporting documents → medium - inception_to_loss_mismatch — loss date precedes policy inception → critical (auto-decline candidate)

2. LLM narrative agent. Reads the FNOL text + claimant statements + adjuster notes. Flags signals rules can't catch: timing inconsistencies, suspicious phrasing patterns, witness contradictions versus the FNOL on file. Tuned to refer under uncertainty rather than accuse — bad-faith mitigation by design.

3. External database lookups. - ISO ClaimSearch (Verisk) — corroborate the claimant's stated loss history against the industry-wide claim database. Discrepancies between stated and actual loss history are a top-tier fraud signal. - NICB Forewarn — VIN lookups (auto claims), claimant name searches against the NICB watchlist, prior-fraud-indicator flags.

The aggregate score

Each fired signal carries a severity (low / medium / high / critical). The aggregate fraud_score (0–100) is severity-weighted: critical = 30 points, high = 18, medium = 8, low = 3. Scores cap at 100.

Thresholds in production: - ≥ 60 → auto-route to SIU queue - ≥ 40 → adjuster sees signals but proceeds with claim - < 40 → standard adjudication

What AI fraud detection does NOT do

• Auto-deny claims. Bad-faith exposure under most state UCSPAs is severe. AI flags; humans decide.
• Replace the adjuster's judgement. A fraud_score of 75 doesn't mean fraud — it means signals are present. The SIU investigator decides.
• Cross-tenant pattern detection. Carriers can't share claim-level data without consortium agreements. Within-book pattern detection is fine; cross-book requires explicit data-sharing frameworks (e.g. ISO ClaimSearch).

Audit trail requirements

Every fraud signal must log: - The rule that fired (with exact threshold values) - The evidence (which fields on the claim triggered it) - The agent that detected it (rule_engine vs llm_agent vs iso_claimsearch vs nicb) - The timestamp + trace id

This is the audit trail that survives both a state DOI bad-faith complaint and a NAIC market-conduct exam. Without it, the carrier is defending the denial in court with no provenance.