American Express published its Intent Intelligence framework on Oct. 13, 2026, proposing a way to capture and authenticate a customer’s instructions before an AI agent buys something. For commerce teams, the announcement moves a practical question into focus: what exactly did the customer authorize, and can that instruction be checked throughout the purchase?

This analysis separates Amex’s announced approach from operational questions a merchant should investigate. It is not a technical implementation guide, a finding that the framework prevents every error or advice about a customer’s legal rights.

What Amex announced

According to Amex’s announcement, the Create Intent API captures purchase instructions and constraints, such as a merchant or maximum budget. The customer authenticates and approves the captured intent. An Intent ID and Proof of Intent Token then represent the authorized instructions and their high-level limits.

Amex says the token can be used to obtain a digital payment credential for checkout instead of exposing the customer’s card number. It describes the intent records as potential controls for later transaction steps, including authorization and dispute handling. These are the company’s descriptions of its framework, available through the ACE Developer Kit, not independent evidence of its effectiveness.

TechCrunch separately reported the framework’s introduction. Amex’s stated ambition to advance a common approach should not be confused with confirmation that all merchants, issuers or agent providers have adopted it.

The merchant’s product information still matters

There is a related preparation effort. In an Oct. 6 playbook announcement, Amex recommended that businesses keep product information current, inspect the purchase journey for inconsistent information and ask their commerce, technology and payment providers about AI-assisted shopping. It advised beginning with a focused improvement rather than attempting every capability at once.

Analysis: A recorded instruction and an accurate catalog solve different problems. A customer could clearly authorize a particular item while the merchant’s listing leaves its size, availability or total cost ambiguous. Capturing permission does not make inconsistent product information correct. A merchant’s evaluation should examine both the instruction and the information used to fulfill it.

A useful test could begin with a small set of products whose attributes are unambiguous. Reviewers could specify the intended item, acceptable variants and maximum total before comparing that request with the proposed purchase. The aim would be to identify where an agent’s interpretation diverges from the customer’s stated instruction, not to count every completed checkout as a success.

Consent needs a meaningful review point

Analysis: The approval moment deserves its own experience test. A customer should be able to understand the proposed purchase without reconstructing a long conversation. A team evaluating an agent could inspect whether the review makes the item, quantity and relevant limits legible. Those are testing questions, not assertions about the exact interface Amex requires.

Exceptions also need attention. What happens if the requested item is unavailable, the price changes or a substitute appears? A merchant should ask its providers which changes require renewed approval and how the decision is recorded. The answer should come from the applicable service specifications, rather than an assumption that a token can resolve every ambiguous request.

Separating these questions helps avoid a false choice between effortless checkout and customer control. A successful evaluation could show that some tasks are suitable for delegation while others need another confirmation. Stopping to clarify an instruction may be the correct outcome, even when it adds a step to the experience.

Measure the completed purchase, not just checkout speed

Analysis: An operational trial should follow the purchase beyond the moment payment succeeds. Reviewers could compare the approved request with the order record, fulfillment and customer support handoff. They should document which records are available to each participant and which questions require escalation.

That does not mean a merchant should invent its own payment protocol. It means the commerce team needs a clear account of its providers’ actual capabilities, responsibilities and limitations before expanding a pilot. Any claim about protection from loss or liability requires the applicable terms and appropriate expert review; it should not be inferred from a framework announcement.

Amex’s release supplies a current reason to ask those questions. The business test remains whether a bounded purchase workflow preserves the customer’s instruction in practice. Faster checkout is only one possible outcome; a purchase that matches the authorized request is the more important measure.