Compliance work sounds like a back-office function, but at Stripe's scale it is a survival condition. The company moves $1.4 trillion in payments annually across 50 countries, which means its compliance teams face a daily flood of transaction reviews, fraud checks, and anti-money-laundering assessments. Miss one, and regulators notice.
The breaking point was not volume alone. It was the way analysts were spending their time. Up to 80% of each review was consumed by navigating fragmented internal systems to gather documentation, not by actually assessing risk. This is a well-known pattern across financial services globally. A recent industry survey found that compliance teams spend as much as 42% of their budgets on false positives and manual reviews. The global cost of financial crime compliance now exceeds $200 billion per year.
Stripe's solution was to build AI agents that gather all the relevant information before a human reviewer opens a case. Think of it like a research assistant who reads every file, pulls every relevant document, and writes up a summary before the analyst sits down. The analyst still reads, judges, and signs off. The AI never makes a call.
This human-first design was not just philosophically cautious. It was the entire reason the system could be deployed in a regulated environment at all. Financial regulators require full audit trails for every compliance decision. Stripe built logging into the system so every step the AI took, every document it pulled, every conclusion it offered, is permanently recorded and retrievable. As one financial services governance guide puts it, regulators treat missing decision traces as a books-and-records violation.
The architecture works by breaking large, complex reviews into smaller chunks. Rather than asking one AI agent to complete an entire investigation, Stripe decomposed the work into bite-sized sub-tasks. Each sub-task feeds into the next, like a checklist where the answers to early questions shape what gets asked later. This keeps each AI task focused and testable. If one step is wrong, it is easy to catch.
The early results: 26% less time spent on each review, and over 96% helpfulness ratings from the analysts using the system. The system now runs more than 100 individual AI agents internally, up from a handful at launch less than a year ago.
This story matters well beyond payments. The same compliance burden pattern applies across insurance, asset management, trade finance, and any regulated industry where skilled professionals spend most of their day doing administrative retrieval work instead of exercising judgment. Research from Experian suggests up to one-third of compliance tasks across financial services could be automated, returning 8 to 12 hours per week to analysts for higher-value work.
Three lessons transfer directly to any organization considering this path.
First, start with tasks that can be pre-completed before a human opens the case. Pre-fetching information is lower risk than real-time AI decision support, and it produces measurable time savings immediately.
Second, keep humans as the decision-maker, not just a checkpoint. This is not a regulatory formality. It is what allows the system to be trusted and expanded. The moment AI is making final calls in a regulated context, the audit trail problem becomes much harder and the liability exposure grows significantly.
Third, log everything from day one. Regulators across the US, UK, and EU are increasingly moving from guidance to active audits of AI systems in financial services. The organizations that built audit trails early are in a fundamentally better position than those trying to retrofit them later.
The $200 billion annual compliance burden across financial services is not disappearing. Regulatory pressure is actually increasing, with compliance demands rising up to 35% year-over-year in some European markets. AI agents that handle data gathering, not decision-making, are how organizations absorb that growth without a proportional headcount increase. Stripe just showed the math works.