Mapfre USA overhauled its insurance fraud detection system using a modern data platform on AWS, incorporating graph-based features and scalable serverless processing. This approach replaced traditional rule-based methods with advanced machine learning models, boosting fraud identification precision and generating multi-million-dollar savings.
- Serverless EMR delivers cost-effective, elastic batch and scoring workloads
- Graph-based features enhance ML fraud detection beyond traditional methods
- Tight integration between data platform and Guidewire Claims drives operational impact
Infrastructure signal
Mapfre USA’s modernization relies on a layered lakehouse architecture featuring Apache Iceberg tables on Amazon S3 with metadata managed by AWS Glue Data Catalog and governed via Lake Formation. This structure supports long-term data governance and agility for evolving business requirements. Compute workloads run on Amazon EMR Serverless, providing on-demand, scalable processing that balances cost with performance. Task orchestration is centralized using Apache Airflow through Amazon MWAA, which ensures consistent job scheduling, monitoring, and fault recovery.
Integration with Neo4j graph database enables advanced fraud feature extraction by analyzing complex networks of claims, policyholders, vehicles, and providers. This architecture enables the platform to incorporate both traditional structured data and graph-enriched insights, improving detection accuracy. Automation in deployment includes environment-based CI/CD promotion and secret management, enhancing operational reliability and security for sensitive insurance data.
Developer impact
Developers benefit from a unified, programmable data environment structured around repeatable workflows orchestrated via Airflow operators, which simplify management of diverse batch pipeline stages. Working with EMR Serverless abstracts the need for fixed cluster sizing, enabling developers to focus on pipeline logic and feature engineering rather than infrastructure tuning. The graph database connectivity allows data scientists to create sophisticated network-based fraud indicators, offering richer model inputs and interpretability.
Close coupling with Guidewire Claims system necessitates robust APIs and retry mechanisms, ensuring ML predictions directly influence claim investigation activities in real time. This reduces manual intervention while providing clear explanations of fraud risk drivers for investigators. The CI/CD pipeline and controlled access policies enable teams to promote changes safely through dev/test/production environments, improving agility and reducing downtime.
What teams should watch
Security and data governance teams must monitor the extensive use of AWS Lake Formation to enforce consistent access controls across datasets that include sensitive policy and claims information. Ensuring compliance with insurance data regulations depends on properly scoped privileges and audit logging within the lakehouse environment. Operational teams should track cost and performance metrics from EMR Serverless workloads, adjusting pipeline scheduling or resource configurations as usage patterns evolve.
Fraud detection and analytics groups should focus on expanding graph-based feature sets to capture emerging fraud trends and validate model effectiveness continuously. Collaboration with claims handling units is crucial to refine the feedback loop between machine-driven fraud flags and manual investigation outcomes. Platform teams need to maintain instrumentation and alerting on Airflow orchestrations and downstream Guidewire integrations to minimize disruptions and enable quick issue resolution.