AI agents are buying things on your behalf now. That’s the pitch from every major AI platform pushing agentic workflows, and it sounds convenient right up until an agent buys Adidas sneakers when you asked for Nike. American Express has introduced a framework specifically designed to stop that from happening, and it’s positioning the approach as a potential industry standard.
The framework is called the Intent Intelligence Framework, and it slots into Amex’s existing ACE Developer Toolkit, which the company first launched in April. The core idea is straightforward: before an AI agent completes any purchase, the customer’s original instructions get captured, authenticated, and locked to a token that travels with the transaction. The agent then uses a secure digital payment credential at checkout, not the actual card number, and that credential is tied to whatever constraints the customer set upfront.
Here’s how it works in practice. A customer tells their AI agent to buy white Nike sneakers in a size 9, under a certain budget. The Create Intent API captures those instructions, including merchant preferences, price limits, and product specs. The customer then authenticates their identity and approves the captured intent. From that, the system generates an Intent ID and a Proof of Intent Token. The token holds the high-level constraints. The agent uses a derived payment credential at checkout. If the agent tries to purchase from a different merchant or exceed the budget, Amex can block the transaction before it goes through.
That last part is the real differentiator, according to Amex VP of Agentic Partnerships and Strategy Hannah Bozian. “Our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow,” she told TechCrunch. Catching a mismatch before the transaction completes is meaningfully different from flagging it in dispute resolution afterward.
The timing matters. AI agents are moving fast from research assistants to active participants in commerce, and the infrastructure to govern that hasn’t kept pace. Large language models are probabilistic by design, which means even well-prompted agents make mistakes. Most current implementations ask users for consent before a purchase, which is fine for now. But Bozian is thinking ahead to a world where agents act more autonomously, reordering household staples, for instance, where the user’s only real constraint is price. In those cases, an automated verification layer isn’t optional, it’s the whole trust model.
The Intent ID and token also have uses beyond blocking bad purchases. They create an auditable record of what was authorized, by which agent, and when. That’s useful for fraud prevention and dispute resolution, two areas where the absence of a clear authorization trail has historically created headaches in card payments.
Amex says it hasn’t seen a spike in disputes from agentic transactions yet. But the company isn’t waiting for that to happen. Survey data from its Trendex report shows 66% of U.S. consumers are likely to use AI agents for price research and 61% for product recommendations in the next 12 months. That’s a lot of potential transaction volume moving through systems that weren’t built with agentic behavior in mind.
The key features of the framework include:
- A Create Intent API that captures purchase instructions and constraints upfront
- Customer identity authentication tied to the captured intent
- A unique Intent ID and Proof of Intent Token generated per transaction
- A secure digital payment credential used at checkout instead of the real card number
- An authorization log for fraud prevention and dispute resolution
Amex is already in discussions with payments ecosystem partners about implementing ACE, including the agentic tokens and other parts of the developer toolkit. The framework is designed to be flexible across different AI models and agentic architectures, which it needs to be if Amex actually wants this to become a standard rather than a proprietary solution. Whether other card networks and payment processors follow with compatible approaches will determine if this becomes infrastructure or just a competitive feature. Either way, the problem it’s solving is real, and it’s only going to get more urgent.
