Stripe cuts compliance review time 26% with AWS AI agents

Stripe has deployed a production AI agent system that cut compliance review handling time by 26 percent while maintaining human decision-making authority and regulatory auditability. The payments processor, which handles $1.4 trillion in annual transaction volume across 50 countries, built the system on Amazon Bedrock to address a scaling challenge: compliance analysts were spending 80 percent of their time gathering documentation across fragmented systems rather than performing risk assessments.

Why it matters

Financial services firms face a $206 billion global compliance burden, and most struggle to scale review operations without proportional headcount growth. Stripe's architecture demonstrates how agentic AI can accelerate compliance workflows without sacrificing the human judgment and audit trails regulators require—a blueprint for regulated industries where automation has historically failed to gain traction.

Architecture: task decomposition and ReAct agents

Stripe rejected the idea of a single agent handling entire compliance reviews. Instead, the company decomposed complex investigations into discrete sub-tasks organized as a directed acyclic graph. Each sub-task can depend on outputs from prior tasks, creating rails that ensure investigations cover required bases while giving agents focused context.

The system uses a ReAct (reasoning and acting) agent framework. For each sub-question, the agent cycles through reasoning steps and tool calls to fetch relevant signals from internal databases and services. Critically, the agent's responses serve only as supplementary research—human reviewers must answer each sub-task themselves. This design preserves accountability while capturing efficiency gains. The system achieved over 96 percent helpfulness ratings from reviewers.

The ReAct implementation includes a closed-loop control mechanism: whenever the agent requests a tool, execution stops, the tool runs programmatically, and its output is injected back as an observation before the agent continues. This pattern grounds reasoning in actual data, prevents hallucination of tool results, and creates an explicit audit trail of every action and rationale.

Infrastructure decisions: dedicated agent service

Stripe initially tried fitting agents into traditional ML inference infrastructure but quickly abandoned the approach. Traditional ML systems are compute-bound, optimized for millisecond GPU-based responses. Agents are network-bound, spending minutes waiting on foundation model calls and tool executions with unpredictable latency.

The company built a dedicated agent service that now hosts over 100 agents, up from a handful at launch less than a year ago. The service handles long-running, stateful interactions through asynchronous execution patterns, allowing threads to manage multiple concurrent sessions without blocking.

Stripe also deployed an LLM Proxy microservice between agents and Amazon Bedrock. The proxy provides a single API across multiple foundation models, implements fallbacks during resource constraints, prevents noisy neighbor problems across teams, and tracks usage for forecasting. Amazon Bedrock's prompt caching capability reduced costs by reusing common prompt prefixes across agent turns rather than reprocessing entire conversation histories.

Lessons from production

Stripe's deployment surfaced several insights. Keeping agent tasks small enough for working memory proved essential—incremental quality testing beats diving into full automation. Asynchronous workflow orchestration with DAG support is necessary for complex agent interactions at scale. And dedicated microservice architecture matters because agents have fundamentally different resource profiles than traditional models.

The system maintains full audit trails meeting regulatory examination standards, with every agent action and decision retrievable historically. As Stripe continues growing, the architecture allows compliance operations to scale proportionally without headcount increases, freeing human reviewers to focus on tougher problems and new investigation opportunities.

These details were first reported by Christopher Phillippi and Chrissie Cui from Stripe in a post on the AWS Machine Learning Blog.