Amex sets out how AI shopping agents should prove what customers asked for

American Express has published a framework setting out how AI shopping agents should capture and prove a customer’s instructions before they buy anything, as card networks and payments firms try to settle who is responsible when an agent makes a mistake.

The framework, called Intent Intelligence, was released on 13 October and details the company’s recommended approach to capturing and authenticating a customer’s purchase intent before an agent completes a transaction. It is designed to work alongside Amex Agent Purchase Protection, a commitment the card issuer announced in April to back cardholders against charges caused by AI agent errors.

How the framework works

According to the American Express announcement, the framework carries a customer’s instructions through a transaction in four stages. A Create Intent API first captures the purchase instructions and constraints, such as the merchant and maximum budget, and the customer authenticates and approves them. An Intent ID and a Proof of Intent Token are then generated, the token is used to obtain a digital payment credential for the agent to use at checkout in place of the actual card number, and both identifiers act as controls through the rest of the purchase to check it matches what the customer approved.

The approach is intended to make that authorised intent available to the card issuer earlier in the transaction, before an agent executes a purchase, rather than after the event. American Express said the framework would support fraud prevention and dispute resolution.

“Building trust into AI-powered transactions is critical to the adoption of agentic commerce,” said Luke Gebb, head of global innovation at American Express.

The company said it was publishing the framework openly in support of common, interoperable standards for agentic commerce. According to the announcement, that work includes “engaging with EMVCo to support secure card-based agentic transactions and supporting the x402 Foundation’s work to advance standards for internet-native payments”. EMVCo manages the EMV specifications used for card payments worldwide.

American Express told Digital Commerce 360 that Intent Intelligence does not yet have any live integrations. The publication noted that rivals are building their own tools, including Visa’s Intelligent Commerce and Mastercard’s Agent Pay.

The limits of protection

Amex Agent Purchase Protection was unveiled in April alongside the company’s Agentic Commerce Experiences developer kit, The Paypers reported at the time. On its agentic commerce page, American Express says that “in the future, if a Card Member authorizes a registered AI agent to make a purchase, and that agent sends American Express the customer’s authenticated purchase intent, American Express will protect eligible customers from charges related to AI agent error.”

The terms set out there show where the hard cases lie. Protection applies only after the cardholder has initiated a return with the merchant where possible, and claims “may not be available in some cases, including where purchase intent is subjective or non-verifiable (e.g. ‘best’ or ‘really nice’)”, according to the company. American Express says full terms and conditions will be made available in the future.

Consumers have reasons for caution. Digital Commerce 360 cited a survey by NMI in August which found that 58% of respondents worried an agent could buy the wrong item and 60% were concerned about overspending.

‘Intent becomes much more subjective’

Laurence Booth, group chief executive of Trust Payments, a UK-based payments and acquiring group, welcomed the moves but said establishing what a customer actually wanted remained one of the hardest questions for the industry. Booth spent 16 years at American Express before joining Trust Payments in 2021, according to the company.

“American Express’s agent purchase protection and the Intent Intelligence framework it published last week are important steps for agentic commerce. Giving consumers clarity over what happens when an AI agent makes a purchasing error is a key part of building confidence in this new way of transacting,” he said.

“They also bring into focus one of the hardest questions the industry still has to answer, which is intent. When a shopper asks an agent to buy a specific product at a specific price, it is relatively clear what they wanted. When they ask it to find ‘the best’ option, or something that is ‘good value’, intent becomes much more subjective. Establishing what the customer actually authorised in those circumstances is a challenge for the whole payments ecosystem and it is encouraging to see moves towards a common industry approach.”

Booth said UK rules had been built around human shoppers. “In the UK, existing payments and dispute frameworks largely assume that a person is making and authorising the purchase. As agents become a genuine commerce channel, issuers, schemes, acquirers and merchants will need a shared way to capture and evidence what the customer actually asked for. That will make it clearer where responsibility sits when something goes wrong and give merchants and acquirers certainty over how agent-led transactions are handled.”

“Getting this right is what will help turn agentic commerce from discovery and research into a trusted way to pay,” he added.

Trust Payments says it employs more than 400 people across nine offices, holds Visa and Mastercard principal memberships, and is licensed by both the UK Financial Conduct Authority and the Malta Financial Services Authority.