JTSTech Services
← All articles

E-commerce · October 13, 2026 · 6 min read

Amex now keeps a signed note of what the shopper told the AI agent to buy. When the agent gets it wrong, the first stop is still your returns desk.

On 13 October American Express explained how it will check an AI agent's purchase against what the customer actually asked for: the merchant, the budget, the brand and size, captured and signed off before checkout. If the agent tries to buy something else, Amex can block it, and it plans to refund shoppers for agent mistakes. Read the small print, though, and the cardholder is expected to try a return with you first. Which makes your product data part of the evidence, and a vague listing the place where an agent's mistake lands.

Written by Jeremy Souffir Founder, JTS Tech Services

The short version, and the direct answer if you read nothing else. On 13 October American Express detailed what it calls its Intent Intelligence framework. Before an AI agent buys something on an Amex card, the customer's instructions are written down (which merchant, the maximum budget, product preferences such as brand and size), the customer signs in and approves them, and the agent is handed a payment credential tied to that approved intent instead of the real card number. If the agent then tries to buy something that does not match, Amex says it can block the purchase. Amex also plans to protect cardholders from charges caused by agent error. The part that matters for a store owner is in the conditions: customers will generally be expected to try a return with the merchant first, and claims can be refused when the request was too vague to check. Put simply, the card network is about to compare what the shopper asked for with what your store sold. The only facts it can compare against are the ones your product data states plainly.

What did Amex actually announce?

  • Intent is captured before checkout. A Create Intent API records the customer's instructions and constraints, such as merchant name, maximum budget and product preferences. How that intent is collected can vary from one AI assistant to another.
  • The customer signs off. The customer authenticates and approves the captured intent, which produces an Intent ID and a "Proof of Intent" token carrying the high-level constraints.
  • The agent pays with a stand-in. The token is used to obtain a payment credential the agent uses at checkout in place of the actual card number.
  • Mismatches can be stopped. Amex's own example: the customer asks for white Nike sneakers in size 9, the agent tries to buy Adidas, and Amex can block the transaction.
  • It is part of a bigger kit. The intent services sit inside Amex's Agentic Commerce Experiences (ACE) developer kit, launched in April, whose listed partners include OpenAI, Google, Microsoft, Stripe, PayPal and Adyen.
  • There is no flood of disputes yet. Amex told TechCrunch it has not seen an increase in agent-related disputes and is acting early, before volumes grow.

What is live and what is still a plan

According to Amex's own page, the intent, account and payment-credential specifications have been available to developers since 14 April. Agent registration and "cart context" (sharing the contents of the cart for checks and disputes) are still in development. The Agent Purchase Protection for shoppers is marked "coming soon", its full terms are not published, and it is limited to US cards issued by American Express. No obligations for merchants have been announced. Everything below about what this means for your store is our reading of the design, not something Amex has asked merchants to do.

Where does a wrong agent purchase end up?

Amex says eligible agent errors include purchases that deviate from the authenticated intent, and that the protection covers the full amount plus reasonable directly incurred charges. Two conditions shape where the cost actually lands. First, customers generally must try to return the item to the merchant before claiming, where a return is possible. Second, a claim may be denied when the intent is subjective or unverifiable, and Amex gives "best" as its example of a word that cannot be checked.

So there are two ways an agent's mistake reaches you. If the check works at checkout, the purchase is blocked, which is good when the agent really did pick the wrong thing and a lost sale when your product was right but your listing did not say so clearly. If the mistake gets through, the first conversation is a return at your counter, not a claim at the bank. In both cases what decides the outcome is whether the plain facts of the order (brand, size, colour, price, who the seller is) can be read off your product data and matched against the customer's words.

The shopper's approved request on the left, written as a short list of plain facts. Your product listing on the right. The check in the middle can only compare the facts both sides state clearly, so anything that lives only in a photo or a paragraph of description falls through the gap.
The shopper's approved request on the left, written as a short list of plain facts. Your product listing on the right. The check in the middle can only compare the facts both sides state clearly, so anything that lives only in a photo or a paragraph of description falls through the gap.

Amex has not published how its matching works, and we are not going to guess at the mechanics. The principle does not depend on them. Any system that checks "white Nike, size 9, under $150" against an order can only use what the order and the product carry in a form a machine can read. If the brand is only visible in a product photo, the size sits in a free-text description, or the price at checkout jumps past the budget once shipping is added, the check either fails a legitimate sale or lets a wrong one through and leaves you to handle the return.

The two conclusions that both get this wrong

Amex is a small share of our card mix, so this can wait

The card brand is the example, not the lesson. In the same few weeks Mastercard began testing a score for how likely a transaction was started by an AI agent, and six banks including Bank of America, Capital One, ING and NatWest published shared principles for trusted agentic commerce. Every one of these is heading the same way: write down what the customer authorised, then compare it with what was bought. Amex is simply the first to describe the comparison in this much detail.

Amex refunds the shopper, so agent mistakes are not our problem

Merchants clearly think the AI provider should pay: in PYMNTS research reported by FinTech Weekly, 93% of surveyed merchants said the agent provider should bear the loss when an agent makes a wrong purchase. The design Amex has published routes the customer to your returns process first. Expectation and design point in different directions, and until the full terms exist, assume the return comes to you.

What is worth doing this quarter?

  • Put the objective facts in fields, per variant. Brand, size, colour and material should each be a structured attribute on every variant, with a GTIN where you have one, not only words in the title, a photo or the description. These are exactly the constraints a shopper hands an agent.
  • Make the final price predictable. Shoppers will give agents a maximum budget. Shipping, duties and fees that only appear at the last step are how a correct product turns into an over-budget order.
  • Use one merchant name everywhere. The intent records which merchant the customer meant. Your storefront name, your product feed and the name on the card statement should be recognisably the same business.
  • Write the returns policy so a machine can read it. If agent mistakes arrive as returns, the window, the conditions and who pays shipping should be clear on the page and in your feeds, not buried in a PDF.
  • Start counting agent orders. Tag orders and returns that come through AI assistants where your platform lets you, so you can see whether the return rate is different before it becomes a cost you notice at month end.
  • Keep a record of what was sold. The exact variant, price and shipping shown at checkout are your side of any future comparison. Amex's cart-context piece, once it ships, suggests the network will ask for exactly that.
The same comparison when the listing carries its facts in fields: every line of the shopper's request has a field on the product to connect to, so the check has something real to hold on to and far less is left for the returns desk.
The same comparison when the listing carries its facts in fields: every line of the shopper's request has a field on the product to connect to, so the check has something real to hold on to and far less is left for the returns desk.

The genuinely encouraging part

None of this is new work invented for Amex. Clean variant attributes, honest pricing and a readable returns policy are the same things that get a product recommended by ChatGPT, Gemini and Perplexity in the first place. The card network's check simply gives the work a second payoff: data that helps an agent choose you also helps prove the agent chose correctly. And Amex says disputes are not rising yet, so this is the cheap moment to do it.

Where we fit

This is a product data problem, and it is the one our AI Shopping Visibility service exists to fix. We work through your catalog so every variant carries its brand, size, colour, material and identifiers as structured attributes, line up your feed, storefront and policies so prices and returns read the same everywhere, and verify the result against the checks AI agents actually run. It is built on CortexCommerce, our own agentic commerce platform, whose CortexGuard app is live on the Shopify App Store. The outcome is a catalog an agent can choose from with confidence and a card network can check against without guessing, which means fewer blocked sales and fewer returns that should never have happened.

Sources

← Back to all articles

Keep reading

Security

One prompt to a public support bot reached every AI agent in the account. AWS says that is documented behaviour, which is exactly the problem.

Strategy

Google's new Gemini coworker gets its own email address and a line in your company directory. Somebody still has to be its manager.

E-commerce

TikTok will let shoppers buy from your brand without leaving the video, and you stay the seller. The ticket in is the same protocol Google started vetting merchants for the next day.

AI Shopping Visibility

If an agent compared your listing with a shopper's request tomorrow, would the facts line up?

We make every variant in your catalog state its brand, size, colour, price and policies plainly, so AI agents can pick you with confidence and card-network checks can confirm they picked right.

See AI Shopping Visibility