Stripe Cuts Compliance Review Handling Time 26% with AI Agents on Amazon Bedrock

Stripe processes enormous annual payment volume across 50 countries, requiring compliance teams to review thousands of transactions daily to identify and mitigate financial crime risk. As Stripe's footprint grew, skilled compliance analysts were spending up to 80% of their time navigating fragmented systems to gather documentation rather than performing high-value risk assessments, making it hard to scale compliance operations without proportional headcount growth.

Stripe built a production-grade agentic compliance system on Amazon Bedrock, decomposing each complex review into a directed acyclic graph of bite-sized sub-tasks handled by a ReAct (reasoning and acting) agent framework that dynamically calls tools to gather relevant signals. Human reviewers remain the final decision-makers on every sub-task, with agent responses surfaced only as supplementary research. The system runs on a purpose-built agent service and an internal LLM Proxy microservice that provides a single API across multiple foundation models, with noisy-neighbor protection, model fallbacks, monitoring, and prompt caching to control cost.

The agentic system reduced review handling time by 26% while maintaining human oversight and a full auditable trail of every action, decision, and rationale, and agent responses achieved over 96% helpfulness ratings from reviewers. Stripe's dedicated agent service grew from a handful of agents at launch to well over 100 agents in less than a year.

• Decomposing a complex review into a DAG of small, quality-tested sub-tasks kept agents focused and made responses auditable, instead of relying on one unconstrained agent for the whole task.
• Agentic workloads are network-bound and long-running (waiting on model or tool calls), not compute-bound like traditional ML inference, which forced Stripe to build a dedicated agent service rather than reuse its ML inference platform.
• Treating agent output as supplementary evidence for a human reviewer — rather than an autonomous decision — preserved accountability while still capturing efficiency gains, and prompt caching kept multi-turn agent costs manageable.