Stripe explains how AI agents help it scale financial compliance

Stripe has published a detailed breakdown on the AWS blog of how it built a production system of AI agents for financial compliance. It is a rare public case: a large fintech company is disclosing not a concept but a real architecture, with lessons on task decomposition, query costs and the human's place in the decision loop.

Architecture: ReAct and a dedicated service

At the core of the system is the ReAct framework (Reasoning + Acting): the agent alternately formulates its next step and executes it, repeating the cycle until it reaches the goal. The approach has been known for a long time, but Stripe invested significant engineering effort in making it production-ready.

The key infrastructure decision is a dedicated agent service, isolated from the rest of the platform. The isolation solves several problems at once:

• independent scaling of agents without affecting core services
• separate logging of every reasoning step for audits
• flexible access-rights management at the service level
• updating and replacing models without risk to payment processing
• error isolation: an agent failure does not spread to other components

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

Task decomposition and orchestration

Complex compliance cases are split into atomic subtasks: document checking, data reconciliation, risk classification, report generation. The agent does not solve a task in a single call. Each step is isolated, which lowers the probability of error and makes it easier to debug a specific link in the chain.

The orchestration patterns were built with an emphasis on reproducibility. The architecture is not tied to one type of task: the same decomposition principles carry over to KYC verification, AML monitoring, counterparty checks and the preparation of regulatory reporting. Stripe describes this as a deliberate decision: to create a horizontally applicable tool rather than a point solution for a single regulation.

Prompt caching cuts costs

In compliance operations, much of the context is stable and recurs: regulatory guidance, corporate policies, risk assessment templates. Stripe uses prompt caching through the Claude API, so these blocks are cached and are not billed in full on each subsequent call. The result is a multiple-fold reduction in inference cost for the same volume of tasks. With thousands of compliance checks per day, the difference between cached and uncached prompts becomes a significant operating expense. The team calls prompt caching one of its main optimization tools, with no loss of quality.

Humans in the loop: auditability over speed

Financial compliance is a field where a mistake is costly in the most literal sense: fines, loss of license, reputational damage. Stripe deliberately kept human oversight at critical points. The agent formulates a conclusion and proposes an action, and 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 consequential decisions.

This approach provides auditability at scale: every decision can be traced back to a specific step of the agent's reasoning. For a regulated company, this is both best practice and a legal requirement.

What this means

Stripe's breakdown is one of the few public examples of production agents in the regulated financial industry. The architectural decisions, the cost data and the auditability principles are material that could otherwise only be gained through costly experience of one's own. For teams building AI agents in finance, law or medicine, it is a rare and valuable reference point.