Stripe explains how AI agents help scale financial compliance

Stripe published a detailed breakdown on the AWS blog of how the company built a production AI-agent system for financial compliance. This is a rare public case: a major fintech company reveals not a concept, but an actual architecture — with lessons on task decomposition, request costs, and the role of humans in the decision-making cycle.

Architecture: ReAct and a Dedicated Service

At the core of the system is the ReAct framework (Reasoning + Acting): the agent alternately formulates the next step and executes it, repeating the cycle until the goal is achieved. This is a long-known approach, but Stripe invested significant engineering effort in its industrial implementation. The key infrastructure solution is a dedicated agent service, isolated from the rest of the platform. Isolation solves multiple problems at once:

• independent scaling of agents without affecting core services
• separate logging of each reasoning step for audit
• flexible access control management at the service level
• model updates and replacements without risk to payment processing
• error isolation: an agent failure does not propagate to other components

The dedicated architecture also simplifies onboarding new compliance scenarios: adding a new type of task does not require changes to the system core.

Task Decomposition and Orchestration

Complex compliance cases are broken down into atomic subtasks: document verification, data reconciliation, risk classification, report generation. The agent does not solve a task in one call — each step is isolated, which reduces the likelihood of error and simplifies debugging of a specific link in the chain. Orchestration patterns were built with an emphasis on reproducibility. The architecture is not tied to one type of task: the same decomposition principles transfer to KYC verification, AML monitoring, counterparty checks, and regulatory reporting preparation. Stripe describes this as an intentional decision — to create a horizontally applicable tool, not a point solution for one regulation.

Prompt Caching Reduces Costs

In compliance operations, many contexts are stable and recurring: regulatory guidelines, corporate policies, risk assessment templates. Stripe uses prompt caching through the Claude API — these blocks are cached and are not charged on each subsequent call. The result is a multiple reduction in inference costs with the same volume of tasks. With thousands of compliance checks per day, the difference between cached and uncached prompts becomes a significant operational expense. The team calls prompt caching one of the main optimization tools without loss of quality.

Humans in the Loop: Auditability Over Speed

Financial compliance is an area where an error costs money in the literal sense: fines, license loss, reputational damage. Stripe deliberately kept human control at critical points. The agent formulates a conclusion and proposes an action — a human confirms or rejects it. The goal is not to remove specialists from the process, but to redirect their attention from routine processing to making significant decisions. This approach ensures auditability at scale: each decision can be traced back to a specific step of the agent's reasoning. For a regulated company, this is both a best practice and a legal requirement.

What This Means

Stripe's breakdown is one of the few public examples of production agents in regulated financial industry. Architectural decisions, cost data, auditability principles — material that can otherwise only be obtained through your own costly experience. For teams building AI agents in finance, law, or medicine, this is a rare and valuable reference point.