An insurance company experiences billions of dol- lars of fraud loss each year. While much of it is detected, there is also a significant amount of undetected fraud, and a considerable operational effort is expended on these fraud investigations. In response, the claims processing operation is transitioning from a rule-based classification system to a machine-learning-driven classification system. The goal of this project is to develop data-driven predictive models that identify fraudulent insurance claims, allowing the company to manage fraud risk more effec- tively and operate more efficiently. Three broad categories of fraud typology are addressed: synthetic fraud, claim-padding fraud, and collusion fraud. In the current environment, this shift enables a more data-driven and factual approach to fraud detection and minimization. In the future, fraud detection may leverage other AI techniques such as transfer learning, causal AI, and advanced modeling techniques on streaming data. The models could evolve into a more sophisticated risk management tool, enhancing the company’s ability to identify fraud attempts in their infancy or assisting in managing fraud risk more holis- tically in cooperation with external vendors. Beyond fraud risk management, they could eventually also support management of other risks within the company, such as operational risk more broadly, risk in underwriting, and risk in business partnerships.