Amex launched a framework that signs an agent's intent before it pays

A team building an AI agent that places orders for a user has a problem that gets harder the more work the agent does. The card the agent charges is the user's, but the decision to spend was the agent's, and nothing in a credit card number says which of the two authorised the latest purchase. American Express released a set of APIs called the Intent Intelligence Framework that moves the authorisation forward in the flow, so a cardholder signs a specific intent before an agent is allowed to spend. TechCrunch covers the launch, which is part of Amex's ACE Developer Toolkit, released earlier in the year.

The flow has four steps. A Create Intent API records what the user told the agent, including the merchant, the budget and any preferences. The cardholder authenticates to Amex and approves the captured intent. Amex issues an Intent ID and a Proof of Intent token. The agent then uses a Secure Digital Payment Credential at checkout, which carries the token instead of the raw card number.

The problem Amex says the framework answers

Hannah Bozian, VP of Agentic Partnerships and Strategy at American Express, told TechCrunch that the design "pulls intent earlier into the transaction flow" so that the risk check happens before the purchase rather than inside the chargeback window after it. In practice the step Amex is adding is a confirmation screen a cardholder sees before an agent commits a transaction, with the agent's reading of what the user wanted written out, and a signed token the merchant can check against the actual purchase.

Amex told TechCrunch the company has not seen an increase in disputes from agentic commerce yet, so the framework is a response to a problem it expects rather than one it is measuring. The two numbers the company included come from its own consumer survey: 66 percent of surveyed U.S. consumers said they were likely to use AI agents for price research, and 61 percent for product recommendations, over the next twelve months. Those are Amex's own survey figures, with no method or sample disclosed in the TechCrunch piece.

Where this lands against the Anthropic and OpenAI patterns

The Intent Intelligence Framework sits alongside two agent commerce patterns a developer has already seen. Anthropic published a commerce agents toolkit tied to Model Context Protocol server shapes, which gives an agent a vocabulary for talking to a storefront. OpenAI has an operator pattern that drives a browser inside a sandbox. Both leave the payment leg to a stored card the user authorised once, and both inherit the problem Amex is addressing: the token the agent carries at checkout does not describe which purchase the user approved.

Amex's framework sits at the network layer rather than the agent layer, so a merchant integrating it works with Amex's existing payment credentials rather than with a particular agent runtime. For a developer this means a check they do not have to build, provided the cardholder's bank is Amex. For other issuers the equivalent does not exist today, so an agent that charges a Visa or a Mastercard still runs on the stored card pattern.

An "intent" is Amex's structured description of a purchase. The framework will succeed to the extent that structure covers the range of purchases a real agent actually tries, and the user confirmation step does not become a drop-off a vendor cannot absorb.

Source

Amex introduces a framework to prevent AI agents from making unauthorized purchases, in TechCrunch.