Amex publishes intent framework for AI shopping agents
American Express published its Intent Intelligence framework on October 13, setting out how a customer’s purchase instructions could reach their card issuer before an AI agent buys anything. The proposal puts authenticated permission earlier in the payment process, where it can inform transaction controls.
Amex had already named Intent Intelligence among the services in its April 14 developer kit announcement. The October publication explains the framework in greater detail. It should be read as another step in that existing initiative.
The practical question is straightforward. When an assistant says it has permission to shop, what can the companies handling the payment actually verify? A conversational request needs to become something that another system can inspect, constrain and later compare with the purchase.
Turning instructions into payment controls
The Amex technical framework describes creating, updating and cancelling authenticated purchase intent. After the user confirms the instructions and selects a card, authentication produces an Intent ID and a Proof of Intent Token. The latter is a JSON Web Token representing the main constraints and can help obtain a scoped payment credential.
Those constraints include a maximum amount, expiry, merchant preferences and a maximum number of transactions. The model can also express dependencies between items, so an additional purchase depends on another being completed. Amex distinguishes delegated instructions from approval of a specific cart, since each creates different opportunities for agent error.
Agents, wallets and vault providers can have different responsibilities across implementations. Issuers receive intent information for risk assessment and credential scoping. The framework also says purchase summaries should omit personal details that are irrelevant to the purchase.
How this fits beside payment mandates
The current AP2 version 0.2 specification provides useful context. It defines Checkout and Payment Mandates, with different parties checking authorization for the goods and for the payment. A trusted interface obtains user consent, while cryptographic links bind the payment permission to the checkout being approved.
AP2 also requires validation and processing rules for each role to run in deterministic code, even when an AI agent participates. Its autonomous mode lets a user approve constraints in advance. The verifier must then check that the completed checkout and payment fit those constraints. Signed receipts preserve evidence of the outcome.
For a team evaluating these approaches, the important comparison is where each check happens. A record of permission is only useful if the component able to stop a purchase receives it in time. Shared vocabulary alone does not establish that two implementations exchange compatible data or enforce the same limits.
Recognizing an agent is another layer
Visa’s Trusted Agent Protocol addresses merchant visibility earlier in the shopping journey. Its documentation describes signatures tied to a merchant, purpose and time window, helping merchants distinguish trusted commerce agents from unwanted automated traffic. It also covers consumer recognition and payment information. Visa labels the product as still undergoing development and deployment, with availability varying by market.
That highlights a useful distinction for merchants. Recognizing the software making a request and deciding whether the person authorized this particular purchase are separate questions. An integration review should trace both, rather than treating a trusted agent identity as unlimited purchasing permission.
What teams should verify before adopting it
A useful acceptance test would start with a permitted purchase, then change one condition at a time. Increase the total, change the merchant, let the authorization expire and cancel the instruction while the agent is still running. Record which system rejects each attempt and what evidence remains for support staff.
Amex describes Intent Intelligence as a core service of its Agentic Commerce Experiences developer kit and says integrations are being developed. The announcement establishes a published framework, rather than evidence that every merchant or cardholder can already use it.
The next meaningful proof will come from implementations that demonstrate these checks across the full purchase journey. Evaluation should include mistakes, retries and cancellations alongside successful transactions. A clean demonstration of buying an item would show only part of the problem that permission controls are meant to address.
