The checkout was built for a person.
On 13 November OpenAI started letting its agents finish purchases for some US users through ChatGPT Wallet. The shopper approves the basket with a fingerprint and the agent pays with a stand-in card number. For anyone who runs a shop, a booking flow or a ticket desk, the question is what happens on your side when the customer arrives as software.
On Friday 13 November OpenAI launched ChatGPT Wallet. Its agents, which it calls dots, can now complete a purchase on a shopper's behalf. The first rollout is narrow: US Pro subscribers over 18, with eligible cards from Visa, Mastercard, American Express or Chase. The agent never sees the real card number. It is handed a substitute number, and the shopper approves the merchant, the items and the total with Face ID or a fingerprint before anything is charged.
Most of the coverage is about the shopper. We are more interested in the other end of the transaction, because that is the end our clients run.
Every step assumed a human
Look at any checkout and count the places it assumes a person is sitting there. The upsell before payment. The delivery slot picker that only makes sense if you can see the calendar. The fraud check that compares typing speed, device and location against what a normal customer looks like. The confirmation email that someone is expected to read. The returns form behind a login.
None of those were wrong. They were designed for the only kind of customer there was. An agent arriving with a tokenised card, a fixed basket and no interest in the page you spent months designing is a different kind of customer, and OpenAI says so plainly: merchants and issuers can still decline, and risk systems may flag legitimate agent purchases.
The first time most businesses meet an agent customer, their fraud rules will decide it is a bot. They will be right, and it will still be a lost sale.
Where it actually breaks
In the systems we build, the payment is rarely the hard part. The hard parts sit either side of it.
- 01Stock and availability. An agent that buys from a page showing stale stock creates an order you cannot fill. A person might notice the item looked odd. Software will not.
- 02Options and edge cases. Size, date, seat, dietary needs, delivery window. If those only exist as visual widgets, the agent guesses, and the guess becomes your support ticket.
- 03Fraud rules. A purchase with no browsing, no hesitation and a fresh card token looks exactly like the thing those rules exist to stop.
- 04After the sale. Refunds, changes and cancellations still assume the customer will log in and click. Who handles the change when the buyer was an agent?
What we would do now
This is a US, invite-level rollout, and nobody in the UK needs to rebuild a checkout this month. But the work that makes a business ready for agent customers is the same work that makes it better for people, so there is little reason to wait.
- 01Make the truth machine-readable. Real stock, real prices and real options in clean structured data, not just on a rendered page.
- 02Separate confirmation from persuasion. Keep the upsell, but make sure the path to a correct order does not depend on it.
- 03Look at your declines. Find out what your fraud rules do with a fast, quiet, tokenised purchase before a customer finds out for you.
- 04Make changes and refunds work without a login screen, through a clear link or reference that a person or their agent can use.
The shopper still approves the basket with their own fingerprint. Everything around that moment was built for a person at a screen, and that is the part worth looking at first.
