If you’ve used ChatGPT or Claude, you know it types an answer back. You ask a question, it gives you text. That’s a conversation. But what if you need AI to do things instead of just talking?
Imagine your business processes millions of transactions every day, across dozens of countries. Some of those transactions need a quick human check to make sure everything is above board. This is the reality for financial platforms like Stripe, which processes a staggering $1.4 trillion in annual payment volume across 50 countries. Every single day, their compliance teams face thousands of reviews.
The big problem? Skilled analysts were spending up to 80% of their time just hunting down information from fragmented systems. They weren’t actually doing the high-value risk assessment they were trained for. This is a common bottleneck, part of a global compliance burden estimated at $206 billion.
Most people might think you could just feed all the transaction data into a giant AI model and ask it, “Is this transaction risky?” That sounds simple. The weird part is, that approach often falls flat. A single, unconstrained AI trying to tackle such a huge, complex question all at once tends to get lost. It focuses on the wrong details, misses crucial context, or simply can’t handle the sheer volume of fragmented information.
Here’s the thing: you don’t just need an AI that talks. You need an AI that acts.
This is where AI agents come in. Think of an AI agent as a smart assistant that doesn’t just respond to your questions, but can actually perform tasks using various tools (specific capabilities, like searching a database, performing a calculation, or drafting an email). It’s not just a chatbot; it’s a do-bot.
Stripe built a system using these agents, and the results are pretty compelling. They managed to reduce the time spent on compliance reviews by 26%. And here’s the kicker: human experts still make all the final decisions, maintaining full oversight and accountability. The AI just makes their job much, much faster.
Stripe’s AI agent system achieved a 26% reduction in review handling time and a 96% agent helpfulness rating, significantly cutting the 80% of time analysts spent gathering data.
What makes an AI agent different for your business?
Traditional automation is great for simple, repetitive tasks. But compliance work often needs judgment, context, and the ability to adapt. That’s where agents shine. Stripe’s approach rests on three pillars:
- Oversight and accountability: Humans stay in charge. The AI helps, but the final decision is always human-made, with configurable approval steps.
- Transparency: Every action the agent takes, every decision it proposes, is fully documented. This creates an audit trail, which is critical for regulated industries.
- Efficiency: Agents handle the heavy lifting of data gathering and preliminary analysis, allowing human experts to focus on the truly complex risk assessments.
How does Stripe actually build these smart assistants? (You can skip this if “how it works” isn’t your thing)
Stripe didn’t just throw a single massive large language model (LLM) – that’s the AI brain like GPT-4 or Claude Opus that generates text – at the problem. Instead, they broke down the complicated compliance review into smaller, manageable pieces called sub-tasks.
Think of preparing a gourmet meal. You wouldn’t just tell a chef, “Make dinner.” You’d break it down: “First, chop the vegetables. Then, sauté them. Next, sear the steak.” Each of these is a sub-task. For Stripe, a sub-task might be “Verify customer identity” or “Check transaction history for anomalies.” These sub-tasks can depend on each other, forming a clear, step-by-step process. This helps the AI agent focus and ensures all the necessary checks are done.
Even with rigorous quality testing, Stripe doesn’t blindly trust the agent’s answer to a sub-task. The agent’s findings are always presented to a human reviewer as supplementary information. The reviewer makes the ultimate call, ensuring that human judgment and accountability are always part of the process.
To actually perform these sub-tasks, Stripe uses a clever design called the ReAct agent framework. ReAct stands for “Reasoning and Acting.” At its core, it lets the AI brain (a foundation model (FM), which is a very large, general-purpose LLM, often accessed through services like Amazon Bedrock) alternate between thinking and doing.
Here’s a simple example of how a ReAct agent works:
The ReAct agent framework cycles through ‘Thought,’ ‘Action,’ and ‘Observation’ steps to systematically solve problems.
Imagine you ask the agent: “What is 10 divided by the number π?”
- Thought: “Do I already know the answer?” No, it doesn’t.
- Action: “I need a calculator.” It calls a calculator tool.
- Observation: The calculator returns the number 3.18309886.
- Thought: “Now do I have the answer?” Yes. It provides the calculation.
For a complex compliance review, this cycle might involve many more steps: Thought: “I need to check the customer’s purchase history.” Action: “Query the CRM database.” Observation: “Here’s the purchase history.” Thought: “I see a suspicious pattern here. I need to cross-reference with known fraud indicators.” And so on.
This “Thought” -> “Tool” -> “Observation” loop is crucial. It forces the agent to ground its reasoning in actual data, preventing it from hallucinating (making things up). It keeps the AI focused and ensures every piece of retrieved information is explicitly processed.
What this means for your business
Stripe’s experience with AI agents offers clear lessons for any business leader. The core takeaway is that for complex, judgment-based tasks, a carefully designed system of focused AI agents, supported by human oversight, can deliver significant efficiency without sacrificing quality or auditability.
- Real efficiency gains: Stripe’s 26% reduction in review handling time translates directly into cost savings and faster operations.
- Scalability: As your business grows (like Stripe supporting millions of companies across 50 countries), AI agents provide a way to scale compliance and other operational tasks without a proportional increase in headcount.
- Auditability: The full audit trail means you know exactly why a decision was proposed, which is vital for regulatory compliance.
- Human empowerment: Instead of replacing humans, these agents free up your skilled employees to do the strategic, high-value work they’re best at.
This isn’t just about financial compliance. This approach applies to any business process that involves complex information gathering, analysis, and decision-making. Think legal document review, customer support troubleshooting, or even complex project planning.
One final note on efficiency: Stripe also learned that prompt caching helps a lot with costs. If an agent asks the same question or performs the same internal thought process frequently, the system stores the answer. The next time, it retrieves the cached answer instead of running the expensive AI model again, saving computational resources.
The future of AI in business isn’t just about bigger, smarter chatbots. It’s about designing intelligent assistants that can reason, act, and work alongside your teams to tackle your most complex challenges. And honestly, that’s a pretty powerful idea.