American Express published a technical framework on October 13 called Intent Intelligence, which records authenticated purchasing instructions before an AI agent transacts. Those instructions can include spending limits and timing constraints.

The published framework is a technical approach, not a change to merchant liability or a compliance requirement. Nothing about it obligates sellers to do anything.

It does raise a question worth asking your payment provider: when an agent orders the wrong thing, what evidence exists about what the buyer authorized?

The Problem It Addresses

An AI agent completing a purchase creates a gap in the record. The buyer gave instructions somewhere, the agent interpreted them, and a transaction appeared. If the result does not match what the buyer wanted, reconstructing the instruction becomes difficult.

That gap matters for disputes. A chargeback normally turns on whether the cardholder authorized the transaction. When an agent executed it, the authorization question splits into two: did the buyer authorize the agent, and did the agent act within what the buyer authorized.

Amex's framework captures the second part before the transaction rather than reconstructing it afterward.

What Sellers Should Ask Their Processor

The useful question is not whether to adopt Amex's framework. It is what your current payment provider retains when an agent-originated order goes wrong.

Ask three things. What record exists of the buyer's instruction to the agent. Whether that record is available to you during a dispute. And how an agent-originated transaction is flagged in your settlement data, if at all.

If your processor cannot answer, that is useful information. It tells you the dispute evidence available on agent orders is whatever you can assemble yourself.

Why This Connects to Catalog Accuracy

The practical failure mode is an agent selecting the wrong variant. A buyer asks for a product, the agent picks a size or color that does not match, and the order arrives wrong.

That is not fraud. It is a fulfillment dispute where the merchant shipped correctly against an order the buyer did not intend.

Agent access to merchant catalogs is expanding. Shopify opened its catalog and Shop Pay to any agent builder this week, which increases the number of interfaces where variant selection happens outside your storefront.

Clean variant data reduces the frequency of that failure. It does not resolve who absorbs the cost when it happens.

Where This Sits Right Now

Agentic commerce infrastructure is being built by payment networks, platforms, and agent developers simultaneously, with different pieces arriving at different times. Amex is addressing authorization evidence. Shopify is addressing catalog access. Neither has addressed what happens when they interact badly.

The gap between those efforts is where merchant exposure currently sits. An agent can reach your products and complete a purchase before anyone has settled how a disputed agent order gets adjudicated.

For now, the available action is knowing what your processor records and watching whether agent-originated disputes appear in your chargeback data as a distinct pattern. Whether that becomes a material cost depends on how much of your volume eventually arrives through agents, which nobody can currently estimate with confidence.