There is a moment in every purchase that we rarely notice: the second before you press the button. You may have browsed half-asleep, skimmed the reviews, let a friend talk you into it. But at the till you wake up. You look at the number. Shops have spent decades shortening that second, with saved cards and one-click buttons and contactless terminals, and none of them ever removed it, because it belongs to you.
The companies building AI agents would now like to hold it on your behalf.
What happened on Friday
On 13 November, OpenAI launched ChatGPT Wallet, a place to keep a payment card that its agents (OpenAI calls them dots) can use to finish a purchase for you. TechCrunch reports that it is open, for now, to a limited group of ChatGPT Pro subscribers in the United States who are 18 or over and hold an eligible Visa, Mastercard or American Express card. It did not arrive as a surprise. In August, RuntimeWire found an unreleased wallet inside OpenAI's Windows desktop app, and for three months it lived as a rumour with a screenshot.
The road here was not straight. In September 2025, OpenAI and Stripe launched Instant Checkout, which let American users buy from Etsy sellers without leaving the chat, together with the Agentic Commerce Protocol, an open standard that gives agents and shops a shared language. A month later PayPal agreed to put its own wallet inside ChatGPT, with the integration due in 2026. By March 2026 OpenAI had scaled Instant Checkout back, moving it into merchants' own apps, according to Digital Commerce 360. And in June, Visa announced that it would bring its payment network, card credentials and fraud systems into OpenAI's products.
So the pieces were assembled in public, one announcement at a time. Friday was the day someone screwed them together.
How an agent pays without holding your card
The plain version is reassuring, and mostly true. When you add a card, the agent never sees the number. The payment network issues a token, a stand-in that can be charged but is useless to anyone who steals it, and the shop receives that instead of your card. Stripe's version, described when Instant Checkout launched, goes a step further: each token is tied to one merchant and one cart value, so it cannot be spent elsewhere or for more.
Then there is consent. At launch, every ChatGPT Wallet purchase shows you the items, the total and the card, and asks you to approve with your face or your fingerprint. That is the second at the till, preserved. But OpenAI has said it plans to add spending limits, so that you can pre-approve purchases the agent then handles on its own. Visa's announcement points the same way, describing payments that run within "spending limits, merchant categories, and required approvals" set by the user.
This is where something quietly shifts. A spending limit is a different kind of consent. Instead of saying yes to a thing, you say yes to a class of things you have not yet seen. We already do this with standing orders and prepaid cards. What is new is handing that yes to software that also chooses what to buy.
Who answers when it buys the wrong thing
Suppose the agent books the wrong flight, or orders the blue jacket when you meant the black one. Who pays?
Nobody has fully settled it. Visa told the Associated Press, in a report carried by Fortune, that it would handle disputes with the two questions it asks of any transaction: did the customer intend the purchase, and did the merchant process it correctly? The hard case is the one where both answers are yes and something still went wrong in between. You approved the agent, the shop delivered exactly what the agent asked for, and the agent misunderstood you. Visa acknowledged that the rules may need to change for that case. OpenAI says the usual card protections, disputes included, still apply. Neither has published a framework for the agent's own mistakes.
I find it steadying to notice how old this problem is. Anyone who has sent someone else to the shops knows it. The difference is scale, and the fact that this errand-runner cannot easily be asked afterwards what it was thinking. There is a newer worry too: a web page could hide instructions meant for the agent, nudging it towards a particular seller or a larger order. The token protects your card number. It does not protect your judgement.
The shops that would rather it didn't
Not every shop wants to be visited by a machine. In July 2025, Modern Retail noticed a line in the robots.txt file of Shopify storefronts (the file a website uses to tell automated visitors what they may do) saying that "buy-for-me" agents, and any checkout that completes payment without a final review step, were not permitted. Shopify's Ilya Grigorik said the line added no new rules, only a pointer to Shopify's own sanctioned route for agents. Amazon goes further: TechCrunch notes that it bans outside AI agents and steers shoppers towards its own assistant. ChatGPT Wallet will not work on sites that block agents.
Part of the resistance is about who stands between a shop and its customer. A merchant that sells through your agent risks becoming a warehouse with a price list, while the relationship, the recommendations and the data move into the chat window. Every protocol in this story is, in part, a negotiation over who gets to keep you.
For now, the second before the button still belongs to you. Your face unlocks each purchase, and any limit will be yours to set. But the whole direction of travel points towards that second becoming a setting rather than a moment: something you decide once, in advance, and then forget.
It may be worth deciding it slowly.
Sources
Visa Partners with OpenAI to Power the Next Generation of AI Commerce, Visa
ChatGPT's new Wallet lets AI agents shop and check out for you, TechCrunch
Shopify quietly sets boundaries for AI agents on merchant sites, Modern Retail
PayPal signs deal with OpenAI to embed payment system into ChatGPT, CBS News

