Stripe's AI agents cut compliance review time by 26% on AWS Bedrock

Stripe, processing $1.4 trillion in annual payment volume across 50 countries, faced a critical compliance scaling challenge. Analysts spent up to 80% of their time navigating fragmented systems instead of performing high-value risk assessments. To solve this, Stripe built a production-grade AI agent system on Amazon Bedrock using a ReAct (Reasoning + Acting) agent framework. The system decomposes complex compliance reviews into composable sub-tasks organized as a directed acyclic graph (DAG), allowing each sub-task to depend on others while avoiding circular dependencies. This architecture reduced review handling time by 26% and achieved over 96% helpfulness ratings from human reviewers, with final decisions always left to human experts.

The infrastructure includes a dedicated agent service that orchestrates the DAG workflow, leverages prompt caching for cost optimization, and maintains immutable audit trails for regulatory accountability. The system allows human-in-the-loop validation with configurable approval workflows. Key lessons include the importance of task decomposition rather than using a single unconstrained agent, the need for transparent orchestration patterns, and the value of separating pre-investigation from dynamic analysis. The result is a scalable compliance engine that maintains auditability and precision while helping identify 95% of card-testing attacks in real time and reducing customer friction by 20%.