Cutting Insurance Fraud False Positives by 50%

Fraud investigation teams have a paradox: their rules-based detection systems are too good at flagging claims, and not good enough at distinguishing real fraud from noise. Investigators spend the majority of their time clearing legitimate claims.

The false positive problem

Traditional rules engines work by matching patterns: claim amount over X and provider Y and region Z. These rules catch fraud, but they also catch thousands of legitimate claims with similar characteristics.

Batch inference plus agentic investigation

The solution combines Databricks ML with AI Functions and multi-agent orchestration:

1. Anomaly Detection: Databricks ML models trained on historical fraud patterns score claims. Batch inference via ai_query in SQL processes bulk claims volumes efficiently.
2. AI Classification: ai_classify categorizes flagged claims by fraud type and severity.
3. Evidence Gathering Agent: For flagged claims, cross-references public records, claims history, and provider billing patterns via MCP tools. Builds a sourced evidence package.
4. Report Agent: Uses ai_summarize and ai_gen to produce structured investigation briefs with evidence chains for the SIU team.

Agent Bricks Multi-Agent Supervisor coordinates the detection-to-investigation pipeline.

The Lakeflow data engineering stack

• Lakeflow Connect ingests claims data from the claims system via managed connectors
• Spark Declarative Pipelines handle external data ETL (public records, watchlists) with streaming tables and materialized views
• Lakeflow Jobs orchestrate batch scoring and agent workflows on schedule or event triggers

All models via AI Gateway

Every model needed for this use case, whether for classification, summarization, or complex reasoning, is available through Databricks AI Gateway on Model Serving. No need to set up external model access. AI Gateway handles rate limiting, payload logging, AI guardrails (safety filtering, PII detection), and usage tracking.

Unity Catalog as the governance backbone

Unity Catalog governs everything: sensitive claimant data with column-level masking, ML models and their lineage, AI functions used by agents, serving endpoints, and the complete audit trail. It is the foundational layer for the entire platform.

Results

Detection rates improve by 29% because ML models catch patterns that rules miss. False positives drop by 50% because statistical scoring is more nuanced than threshold-based rules. Investigation time decreases by 60% because the Evidence Gathering Agent does the data collection that investigators used to do manually.

Every step is logged via AI Gateway and traced via MLflow, creating defensible evidence chains that hold up in litigation.