Why Rule-Based Fraud Detection Fails
The insurer's existing system used 47 hand-crafted rules: "Flag claims over $15,000," "Alert on 3+ claims from the same address in 12 months," "Review claims with same-day accident and treatment." These rules were easy to understand — but fraudsters had learned to route around them, and the rules generated 8 false positives for every actual fraud case, causing alert fatigue.
The Multi-Model Architecture
We built three specialized models that work in concert: (1) An XGBoost claim-level scorer that evaluates 140 features of each individual claim — amount, timing, claimant history, repair shop reputation, medical provider billing patterns. (2) A Graph Neural Network that maps relationships between claimants, attorneys, repair shops, and medical providers to detect fraud rings — groups of entities that appear in multiple claims together. (3) A text analysis model that reads adjuster notes and medical reports to identify inconsistencies between the narrative and the claimed damages.
The Investigation Brief Feature
The most impactful feature wasn't the fraud score itself — it was the AI-generated investigation brief. When a claim is flagged, the system automatically generates a 1-page summary explaining why the claim was flagged, what specific anomalies were detected, which other claims share network connections, and what evidence to request. SIU investigators reported this reduced their investigation setup time from 4 hours to 45 minutes per case.
Results and Ongoing Learning
In the first year, the system prevented $3.8M in fraudulent payouts — a 12x return on the implementation investment. The model continuously learns from confirmed fraud and legitimate claim outcomes, improving its accuracy over time. False positive rate dropped from 72% to 18% within 6 months as the model learned the specific fraud patterns in this insurer's book of business.