MAPFRE's Fraud Detection Jumped 135% Accuracy — Using EMR Serverless and Graph Analytics
Cloud Platform EngineeringRules-based fraud detection has a fundamental problem: the rules are visible. Sophisticated fraud rings adapt around them. MAPFRE USA’s Atenea data platform replaced that approach with a graph-enriched machine learning pipeline on AWS, and the results shift how the economics of AI in insurance look: 50–135% accuracy improvement over baseline methods, and $6.81 million in realized savings in 2025 — exceeding projections by over $500,000 in the pilot phase.
The Architecture
The previous system relied on historical claim patterns and structured-data-only analysis — which couldn’t detect hidden relationships between claimants, policies, vehicles, and providers that indicate coordinated fraud rings. The Atenea platform addresses that by layering graph analytics on top of the ML pipeline.
Neo4j provides the graph layer, generating 54 graph-based features including suspicious claim linkages and provider fraud ratios. These features expose network-level patterns that a claims adjuster reviewing individual records would never surface: a provider appearing across twenty claims in a pattern that looks random until visualized as a graph, or a cluster of claimants with overlapping vehicle histories and shared phone numbers.
The compute layer runs on Amazon EMR Serverless for elastic batch processing and real-time scoring — no cluster management, no idle compute cost during low-volume periods. Storage follows a three-layer lakehouse model built on Apache Iceberg tables in Amazon S3: Silver (source data), Gold (intermediate processed), and Platinum (feature store with encoded features and predictions ready for scoring).
The integration path to the business system was automated: AWS Lambda functions push model outputs directly into Guidewire Claims, creating adjuster activities without requiring manual pipeline monitoring.
Why Graph Matters Here
Standard fraud detection approaches treat each claim as an independent event. Graph analytics treat the dataset as a network and ask which entities are connected, how often they appear together, and what the structure of those connections looks like compared to known fraud patterns.
The 54 graph-derived features from Neo4j are doing the work that human investigators would spend weeks doing manually — tracing relationships across large claim volumes to find the statistical anomalies that indicate coordinated fraud. That analysis, automated and running on every new claim at scoring time, is why the accuracy jump over baseline methods is as large as it is.
This architecture sits at the intersection of AI agents in production and managed knowledge base patterns that are becoming standard for enterprise AI workloads — the same principles of tiered data, automated inference, and business-system integration apply across verticals.
Business Results
The numbers from MAPFRE USA’s 2025 production deployment:
- MA Auto (Massachusetts Auto): $6.59M in savings
- MA Home: $225K in savings
- Combined 2025 savings: $6.81M
- Net present value at approval: $4.7M
- Pilot phase: Exceeded savings targets by $500K+
The accuracy range — 50% to 135% improvement depending on fraud type and detection method — reflects that different fraud patterns respond differently to graph enrichment. The floor is still a 50% gain over prior methods, which is substantial for a production fraud detection system.
The So What
MAPFRE’s case is one of the cleaner examples of graph analytics producing ROI at a scale that justifies the infrastructure complexity. The EMR Serverless choice keeps the compute elastic without requiring a dedicated Spark team to manage cluster tuning. The Neo4j graph layer adds the relationship intelligence that rules-based systems can’t provide. And the Lakehouse architecture on S3 Iceberg gives the data science team a versioned, queryable feature store that doesn’t require rebuilding with every model iteration.
For insurance, financial services, or any sector where fraud rings exploit relationship structures that claim-level analysis misses, the architecture pattern is replicable. The investment in graph enrichment pays off fastest when the underlying fraud is coordinated — which is precisely where the most expensive fraud concentrates.
Source: AWS Architecture Blog
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →