American Express publishes a purchase-intent framework for AI agents

Sourced news briefing

American Express published its Intent Intelligence framework on 13 October 2026. Its announcement describes capturing and authenticating customer purchase instructions before an AI agent transacts, then preserving that intent through the transaction lifecycle. The framework is available through its ACE Developer Kit. This is a proposed approach to trusted agentic commerce, not a guarantee against disputes. The independent analysis below explores the merchant records and operating decisions that become important when software acts for a buyer.

Independent merchant analysis: distinguish a request from authority

A shopping instruction can contain several layers of meaning. A customer might ask an assistant to find a jacket, compare alternatives and purchase only if the delivered price remains within a limit. The first two actions do not necessarily authorise the third. Merchants preparing for automated buying should distinguish browsing, basket creation and permission to spend. Treating every interaction as equivalent would make it difficult to explain why a particular order was accepted.

The instruction also needs boundaries. Product characteristics, maximum amount, delivery destination and purchase deadline can each affect whether the eventual order matches the customer's intention. These examples are a merchant design exercise, not a description of mandatory framework fields. The practical point is that a general instruction to shop should not silently become unrestricted permission to select any product, quantity or delivery service that the agent encounters.

For a hypothetical travel accessories seller, a buyer may authorise one carry bag under a stated total price. An agent could find a bag within that price but add expensive delivery or a recurring membership. The merchant should be able to distinguish the approved item from additional commitments. Checkout design needs to make material changes visible and, where required by the agreed process, obtain renewed customer approval before creating the final obligation.

Independent merchant analysis: connect permission to the final order

The permission record and the order record should be connected through a reliable reference. A support team needs to understand which instruction applied to which purchase, including the version in effect when the order was submitted. Storing a generic consent notice separately from the final basket is less useful. The record should help explain the transaction without requiring staff to reconstruct a conversation from unrelated logs.

Changes after authorisation deserve particular attention. Stock may disappear, delivery charges may change or a promotion may expire. The merchant should define when the order can proceed, when the agent may propose an alternative and when the customer must decide again. These decisions should be explicit in the integration design. A broad objective such as completing a purchase efficiently cannot resolve every material change on the customer's behalf.

Time also matters. An instruction given during one shopping session may no longer reflect the buyer's needs days later. Merchants should ask how expiry, cancellation and replacement instructions reach their systems. If a customer withdraws permission after an agent has submitted an order, staff need to know which action occurred first and whether fulfilment can still be stopped. A coherent event history supports resolution without promising that every cancellation can be reversed.

Independent merchant analysis: keep identity and intent separate

Evidence that an agent is recognised does not by itself explain what the buyer authorised. Conversely, a detailed shopping request does not establish that the requester was entitled to use a payment credential. Merchant diligence should examine both questions and understand which participant supplies each signal. Avoid combining several different checks into a single label such as trusted transaction that obscures the basis for accepting an order.

Account access creates another boundary. An assistant may need product information or order status without needing broad access to the customer's account. Merchants should ask what access is necessary for each task and how it is limited. The same principle applies to staff investigating an automated purchase. Support should retrieve relevant evidence through authorised tools rather than exporting entire conversations or collecting extra personal information without a defined purpose.

Consider a shared household account in which one person researches a product and another person pays. The merchant needs a practical way to identify the customer relationship and authorised purchase without assuming that every account user has identical authority. This example illustrates an implementation question rather than prescribing a legal answer. The actual solution depends on the service design, the participating providers and the applicable customer terms.

Independent merchant analysis: prepare for imperfect outcomes

Automated buying will still produce ordinary retail problems. A correctly authorised order may arrive damaged, contain the wrong item or disappoint the customer. An intent record helps explain the requested purchase but does not substitute for delivery evidence, product descriptions or customer service. Merchants should preserve these records together so staff can distinguish a dispute about permission from a complaint about fulfilment or quality.

Returns can be particularly awkward when an agent selected a product using conditions that the merchant never displayed clearly. Product attributes, compatibility and cancellation restrictions should be structured consistently across the catalogue and checkout. If the merchant changes those details, keep the version associated with the order. This is useful even without automated shopping, but delegation makes the connection between advertised attributes and accepted instructions especially important.

A merchant should also decide how its support team communicates with an agent and with the customer directly. Some situations can be resolved through order status messages; others require the buyer's decision. Define which requests the assistant can make, such as obtaining a receipt, and which need additional authority, such as changing a refund destination. Do not let convenience turn a narrowly authorised purchasing role into unlimited postpurchase control.

Independent merchant analysis: evaluate readiness with realistic tests

A pilot should include more than successful purchases. Test a basket that exceeds the spending limit, an unavailable item, a changed delivery date, an expired instruction and a duplicate submission. Confirm that the order system records an understandable outcome for each. These scenarios assess whether the merchant's integration respects boundaries. They do not assume that any specific provider automatically solves every scenario through a framework alone.

Measure unexplained orders and support resolution time alongside purchase completion. A system that places orders quickly but leaves staff unable to explain them creates downstream work. Review a sample of completed purchases and ask whether someone unfamiliar with the implementation can understand the customer's instruction, the final basket and any changes. If the answer requires engineering interpretation, improve the operational record before increasing transaction volume.

Commercial teams should describe readiness precisely. They may be testing an integration, reviewing a framework or operating a limited pilot; those stages are different from broad production availability. Avoid telling customers that AI purchases are inherently protected or dispute free. Clear communication about capabilities and limitations will make later adoption more credible and help prevent a technology announcement from becoming an unsupported promise in the sales process.

Record who approves changes to these controls. Otherwise a later catalogue update or integration release could alter purchasing boundaries without the relevant business review.