Amex ties AI purchases to approved buying instructions

Approve the instructions before issuing payment credentials. A software connection called the Create Intent API records the merchant, spending ceiling and other preferences. The shopper verifies their identity and signs off on that record. Amex then creates an Intent ID, which identifies the instructions, and a Proof of Intent Token, a digital representation of their main limits. The token can obtain a checkout credential constrained by those instructions. The agent uses that credential instead of receiving the actual card number. A valid card and a valid instruction are different checks; this design brings both into the payment process.

Permission can expire, change or cover several purchases. The technical framework describes services to create, update and cancel approved instructions. The shopper confirms an expiry date. Permission can produce one credential or several, including purchases from different merchants. A maximum transaction count limits how many credentials can be issued. Item dependencies can also be specified: buying insurance, for example, can depend on first buying the underlying flight. These controls matter when an assistant acts later, rather than asking for approval immediately before each checkout. The public framework leaves required fields dependent on the implementation and use case.

The issuer gets an evidence record; the merchant still has an order to fulfill. An agent translates the shopper’s request into structured instructions. Depending on the integration, an intermediary that securely stores tokens or a digital wallet may store and send the approved record. The issuer uses it to assess risk, constrain payment and compare a later complaint with the shopper’s approval. Identity sign-off can use a passkey; the issuer can also permit an agent or wallet to perform that step. None of this establishes a new merchant fulfillment service, return policy or universal rule about who pays for agent mistakes. It establishes a proposed permission and evidence route.

This builds on earlier work, and benefits still need proof. Amex’s April 14 ACE announcement already included Intent Intelligence and described agent-error protection as a future conditional commitment. October supplies more detail rather than inventing the concept. Visa’s April 2025 initiative also described spending conditions and payment signals for disputes. Amex’s framework is therefore part of a wider effort to give AI purchases enforceable boundaries. Amex says agent-related disputes have not increased; partner implementation remains under discussion. There is no disclosed adoption count or measured reduction in disputes to justify calling this a deployed safety breakthrough.