Production-grade AI agents for financial compliance: Lessons from Stripe

This article describes how Stripe built a production-grade AI agent system on Amazon Bedrock for financial compliance that reduced review handling time by 26% while maintaining human oversight and achieving over 96% helpfulness ratings.

• Stripe processes $1.4 trillion annually across 50 countries, requiring compliance teams to review thousands of transactions daily
• Implemented ReAct agent framework with task decomposition into bite-sized sub-tasks as a directed acyclic graph (DAG)
• Built dedicated agent service with LLM Proxy microservice for standardized access to foundation models with prompt caching
• Maintained human reviewers as final decision-makers, using agent responses as supplementary research rather than autonomous decisions
• Achieved 26% reduction in median review time with full audit trails meeting regulatory examination standards
• Key lessons: keep agent tasks small, use async DAG orchestration, build dedicated infrastructure for network-bound agents, and keep humans in control

Stripe demonstrated that agentic AI can significantly improve compliance operations when properly constrained with rails, dedicated infrastructure, and human oversight, with plans to expand to more complex multi-step investigations.