Stripe details agentic AI system for financial compliance built on AWS

Stripe describes building a production-grade AI agent system for financial compliance on AWS using Amazon Bedrock. The system integrates a ReAct agent framework with task decomposition, orchestration, and prompt caching to scale compliance operations while maintaining human oversight.

According to the post, the system reduced review handling time by 26% while achieving over 96% helpfulness ratings, with human experts remaining firmly in control of final decisions. Stripe notes that skilled analysts previously spent up to 80% of their time navigating fragmented systems rather than performing high-value risk assessments.

The architecture decomposes complex reviews into composable sub-tasks arranged as a directed acyclic graph (DAG), with each sub-task verified through quality testing. Human reviewers interact with an orchestrator that pipes reviewed answers as context for subsequent questions, ensuring oversight and accountability.

Stripe implemented a compliance agent using a form of the ReAct (reasoning and acting) framework on Amazon Bedrock. The agent dynamically gathers relevant signals via tool calls, stopping LLM execution to programmatically run tools and force tool outputs as observations before continuing—implementing a closed-loop control mechanism to ground reasoning in actual data and prevent hallucinations.

The post highlights three pillars: oversight and accountability with configurable approval workflows and multi-layered checkpoints; transparency via full audit trails with immutable documentation; and efficiency through pre-investigation and dynamic analysis to enable deeper reviews at faster pace.