Production-grade AI agents for financial compliance: Lessons from Stripe
TL;DR
Stripe and AWS describe an agent system for financial compliance reviews: millions of businesses, 50 countries, about $1.4 trillion in yearly payment volume, and thousands of manual checks per day set the scale. The agent is not used as an autopilot. Reviews are split into small sub-questions, orchestrated as a DAG, and human reviewers still make the final calls. Agent answers provide pre-work and context.
Nauti's Take
Most agent demos sell magic. Stripe sells infrastructure, which is much more useful.
The real progress sits in the boring layer: orchestration, cost tracking, fallbacks, logs, prompt caching, and clear human ownership. Anyone building in compliance, finance, or legal AI should read this less as Bedrock marketing and more as a warning: without process design, an agent is just an expensive chatbot with extra risk.
Briefingshow
The important part is not that Stripe built a compliance agent, but how tightly the agent is constrained. Small tasks, tool traces, human decisions, and full logs make the system closer to a reviewed research assistant than an autonomous judge. For regulated industries, that boundary is the product.
Sources
