Amex launches framework to stop AI agents from making unauthorized purchases

The problem the framework addresses is the nondeterministic nature of large language models, which makes AI agents prone to doing things they were not instructed to do, including completing unauthorized transactions. Amex says there should be a way to ensure a customer's purchase instructions are carried through the entire transaction lifecycle, from the purchase to everything that follows.

"These new transactions will have a consumer, a merchant, and an agent," Hannah Bozian, American Express vice president of agentic partnerships and strategy, told TechCrunch. She said the company wants agentic transactions to remain "seamless, trusted, and secure, the way our card members expect them to be."

The framework extends the Agentic Commerce Experiences developer toolkit, known as ACE, that Amex launched in April. The intent framework is now among the services included in that toolkit.

In practice, the process starts with a Create Intent API, which captures the customer's purchase instructions along with constraints they may have included, such as the merchant's name, a maximum budget and other product preferences. Bozian said how that intent is collected may vary depending on the AI model involved, but the plan is to make the framework flexible enough to support all agentic commerce models. The customer then authenticates their identity and signs off on the intent the agent captured.

An Intent ID, a unique identifier representing the intent, is then created along with a Proof of Intent Token that contains the customer's high-level constraints. That token is used to obtain a secure digital payment credential that the agent presents at checkout instead of the customer's actual credit card number, which Amex says keeps the purchase secure while aligning it with what the customer authorized. Amex said the Intent ID and token can also be used in other parts of the transaction, including fraud prevention and dispute resolution.

"Our approach differs from the implementations of others in that it pulls intent earlier into the transaction flow. And this is important because it helps mitigate some risk before the risk occurs," Bozian said. She described a case in which a customer asks an agent to buy white Nike sneakers in size 9 but the agent submits a request to purchase shoes from Adidas; Amex could block the transaction because the agent made a mistake.

Bozian said that while many agents today ask for customer consent before making a purchase, Amex expects agents to act far more autonomously in some situations. A consumer might, for example, let an agent re-order household staples such as paper towels or toilet paper, categories in which price and size are the only real constraints and merchant or brand do not matter. She said a log of what was authorized, by which agent and when, could also help address fraud committed on behalf of the agent.

Amex said it has not yet seen an increase in disputes tied to agentic commerce, and that it built the technology proactively as that type of commerce scales. In a recent Amex Trendex report, the company said, 66 percent of surveyed U.S. consumers said they are likely to use AI agents to research the best price for a product or service, and 61 percent said they are likely to use one for a product or service recommendation over the next 12 months.

Amex said it is already in discussions with partners across the payments ecosystem about implementing ACE, including the agentic tokens that contain Amex's payment credentials and other parts of the developer kit, and that it plans to keep developing ACE as the agentic commerce ecosystem evolves. Bozian is scheduled to discuss the Intent Intelligence Framework in more detail on October 13 on the Industry Stage at TechCrunch Disrupt 2026.