Specoria
Menu
TR Free test

InsightsAgentic commerce

Who pays when a shopping agent gets it wrong?

American Express says it will cover errors by registered AI agents. How disputes over agent purchases work today, and what evidence a store should keep.

By 8 min read
Two plaster hands hold a paper balance scale, with parcels on one pan and a small paper key on the other, in front of a torn browser frame.
When a machine makes the purchase, the real question is whose shoulders carry the weight. Illustration: Specoria · generated with AI

A shopper asks an AI agent for a pair of running shoes in size 42. The agent orders size 44, or two pairs, or the right shoe from a store the shopper has never heard of. The card is charged. A week later the shopper calls the bank and says: I never wanted this.

Until recently, that scene was hypothetical. It is less so now. Agents from OpenAI, Google, Meta and others can complete purchases, and the card networks have spent the past year building payment rails for them. What has lagged behind is the answer to a simple question: when an agent makes a mistake, who carries the cost?

On 6 October 2026, American Express gave the first partial answer from a card network aimed at merchants. It is worth reading closely, because it settles less than the headlines suggest.

What American Express announced

Amex published the Amex Business Playbook for Agentic Commerce, a free guide for businesses, and announced two related initiatives.

  • Agent Purchase Protection. Amex says it intends to protect eligible card members and merchants from charges tied to errors by registered AI agents. According to the press release, coverage applies when the card member's purchase intent has been authenticated, and it is subject to terms and conditions.
  • A Merchant AI Advisory Council. A pilot forum where merchants can share input on how Amex approaches AI-driven commerce.

The detail that matters most is the word "intends". The terms have not been published, the release does not explain how an agent becomes "registered", and no start date is given. According to Yahoo Finance, the announcement is US-only for now. This is a statement of direction, not a product a store can rely on today.

The direction itself is notable, though. In April, when Amex first said it would protect customers from agent errors, the commitment was framed around the card member. In an analysis published this year, Worldpay's head of agentic commerce, Reshmi Suresh, noted that Amex's protection covers the cardholder, not the merchant. The October announcement names merchants explicitly, and Amex calls it an industry first. For merchants, that is the part to watch: a card network saying in public that the store should not automatically be left holding the loss for an agent's mistake.

The same release includes survey numbers that explain why Amex is moving. In a Teneo survey of 502 US business leaders carried out for Amex in August 2026, 83% said they see agentic commerce as an opportunity, and 86% said they expect to need at least moderate support from payment providers. Among 2,005 US online shoppers, 52% said they were comfortable with an agent buying everyday, low-cost items. Only 15% said the same about high-value items.

How a disputed agent purchase is handled today

A blank paper ledger lies open with a magnifying glass over one row, beside a plaster hand holding a paper key and a sealed envelope with a violet seal.
Only a clear record can show later what the shopper truly asked for. Illustration: Specoria · generated with AI

Outside the Amex announcement, the honest answer is: the same way as any other online card payment.

There is no separate dispute track for agent purchases. Chargeback specialist Chargeflow puts it plainly: a dispute from an agent-placed order is still a card dispute, with the same reason codes and deadlines. What decides the outcome is what the cardholder claims, not who clicked the button.

  • "I never authorised this." The claim is treated as card-not-present fraud (for example, Visa reason code 10.4). If the agent used a properly issued network token, fraud liability generally follows the existing rules for tokenised payments, which tend to put that risk on the card issuer.
  • "This is not what I asked for." The claim becomes a consumer dispute, such as Visa 13.3, "not as described". This is where agent errors usually land, and where the merchant is most exposed.
  • "It never arrived." A non-receipt claim, judged on delivery evidence as before.

The second case is the hard one. As Worldpay's analysis puts it, no liability shift exists yet for agent errors that fall short of fraud. If an agent misreads an instruction and buys the wrong item, network rules give no settled answer on whether the merchant, the issuer, or the company behind the agent should pay. In practice, when a merchant cannot show what the shopper actually approved, issuers tend to side with the cardholder.

Shoppers notice this too. In Checkout.com's 2026 agentic commerce research, 25% of consumers said they would stop using an agent if a purchase was difficult to dispute.

Evidence is moving to the centre

If disputes will be decided on who can prove what the shopper approved, the industry's answer is to make that approval provable.

The clearest example is Verifiable Intent, which Mastercard and Google introduced in March 2026. According to Fintech News Singapore, it creates a tamper-resistant cryptographic record linking a user's identity, the instructions they gave the agent, and the resulting transaction. Mastercard describes it as a shared source of truth for verifying transactions and resolving disputes, and its chief digital officer Pablo Fourez summed up the goal: when something goes wrong, "everyone needs facts". The specification is published as an open draft (v0.1) under the Apache 2.0 licence.

The Amex wording points in the same direction. Its protection applies when purchase intent has been authenticated, which is another way of saying: the agent must be able to show what the shopper agreed to.

None of this is something a typical store installs this month. Verifiable Intent is a draft, Amex's terms are unpublished, and most agent purchases today still pass through ordinary checkouts. But it tells you what evidence will count in a dispute: a record of the shopper's instruction, the agent's identity, and what was actually ordered.

What a store can do now

You do not need to wait for the rules to settle to reduce your exposure. Most of the useful steps are about making the order unambiguous and keeping a record of it.

  • Make the order summary impossible to misread. Variant, size, colour, quantity, unit price, shipping cost and total should all appear as text on the final confirmation step. An agent that has to guess from an image or a colour swatch is an agent that orders the wrong thing.
  • Keep a confirmation step for unusual orders. A large quantity, an unusually high basket, or a mismatch between the delivery and billing country are good triggers for an extra confirmation by email or SMS. Worldpay suggests the same idea: set spending thresholds that bring a human back into the loop.
  • Record what you can about the agent. If a request arrives with a signed agent identity or a known agent user agent, store it with the order. We covered how agents identify themselves earlier; that information is evidence later.
  • Send a clear order confirmation email. It is often the first thing the shopper sees after the agent finishes. A shopper who spots a mistake in the confirmation email will ask for a cancellation, which costs far less than a chargeback.
  • Make the return path the easy path. In many markets, including Turkey and the EU, consumers already have a statutory withdrawal right for most online purchases. A wrong size ordered by an agent should become a return, not a dispute. That only happens if your return policy is easy to find and read, for people and for agents.
  • Ask your payment provider. Ask whether they pass network token and agent data through to you, how they label agent-initiated payments, and what evidence they will need from you in a dispute. Checkout.com's research found that 76% of merchants expect payment providers to play a decisive role here; it is fair to ask what that role looks like.

The bottom line

American Express has said out loud what the industry had been avoiding: agent errors will happen, and someone has to pay for them. Its answer is not live yet, and it is limited to registered agents and authenticated intent. Until the terms arrive and the other networks follow, a disputed agent purchase is just a card dispute, and the store with the clearest record of what was ordered and confirmed is the store most likely to keep the sale.

Clear order summaries, readable return terms and checkout steps an agent can follow without guessing are also what decide whether an agent completes the purchase in the first place. Specoria sends real agents through shopping tasks on your store and shows where they hesitate or pick the wrong option. Start with a free readiness test.

Sources