Amex wants to prove what you told your AI to buy, before it buys

If you let an AI agent buy something and it orders the wrong size, who eats the cost, and who can prove what you actually asked for? American Express is betting the answer should be written down before the purchase happens. As Digital Commerce 360 reports, Amex has published the technical details of its Intent Intelligence framework, which records a shopper's instructions so card issuers can later check whether an agent followed them. The part worth paying attention to is not the capability itself but what Amex did with it: it gave the recipe away.

Publishing a spec is the strategy, not the product

Amex confirmed to Digital Commerce 360 that Intent Intelligence has no live integrations. Nothing ships here. What shipped is a document other payment and technology firms can build on, and that distinction matters more than it looks. In agentic commerce, the company that defines how intent gets captured and passed between an agent, a wallet and an issuer sets the terms everyone else works within. Amex is trying to make its approach the common one before Visa, Mastercard, Google or OpenAI's protocols lock in theirs.

The mechanics are sensible. A shopper approves what to buy, where, and up to how much; the system mints an identifier and a token carrying those limits; the agent uses a scoped payment credential at checkout instead of the real card number. The useful idea is timing. The authorized intent reaches the issuer before the agent transacts, not after a dispute lands. That gives issuers something to match a purchase request against in real time, and gives merchants a record of what the customer actually signed off on.

The catch is everyone else has a standard too

Here is the problem a published framework doesn't solve on its own. Visa has Intelligent Commerce, Mastercard has Agent Pay, OpenAI and Stripe built the Agentic Commerce Protocol, and Google has both AP2 and a Universal Commerce Protocol already live with retailers like Newegg, Walmart and Macy's. A merchant deciding how to handle agent purchases is not choosing one standard. It is staring at several, each backed by a large company that would prefer its own to win.

Amex says Intent Intelligence is meant to work regardless of who collects the shopper's instructions, whether they go straight to the issuer or through a wallet. That is the right ambition. But interoperability claims are cheap until a second company actually implements them. Until an agent or wallet provider ships a working integration, the fraud detection, authorization and dispute-resolution benefits stay on the slide deck. So does Agent Purchase Protection, the chargeback coverage Amex floated in April, which it confirmed is not available because it requires exactly the integrations that don't yet exist.

Why retailers should care before it's real

Consumer nervousness is the demand side here. In an August NMI survey of 1,000 U.S. adults, 58% worried an agent would buy the wrong item and 60% feared it would overspend. Tellingly, 62% said they'd blame the AI provider for a premature purchase, 34% the retailer and 24% the payment provider. A documented intent record is really a liability-assignment tool: it decides whether the agent, the merchant or nobody is at fault.

For retailers that is worth tracking even now, because the format that wins determines how much evidence you'll have in a dispute and how much engineering you'll spend supporting rival schemes. The honest read is that Amex has made a credible opening move in a standards race it cannot win alone. The next adopter announcement, not this publication, is the thing that tells you whether it's working.