My Agent Got a Credit Card. I Still Have to Face-ID the Sweater.

OpenAI's new ChatGPT Wallet lets AI agents pay. Today, every purchase goes through your thumb. Later, OpenAI says, you might set limits in advance and let the agent shop without asking every time. A stand-up reading of the fine print.

A card reader at a supermarket register, with a card slot and a receipt slot
Photo: Syced, "LaCuCa card reader – payment terminal at supermarket register", via Wikimedia Commons, CC0 1.0

Your agent got a credit card.

You still have to Face-ID the sweater.

That's the whole product in two lines, and honestly, I've seen looser governance from humans with corporate cards.

On Friday, 13 November 2026 — yes, Friday the 13th, the day you'd naturally pick to hand a robot your money — OpenAI announced ChatGPT Wallet. In OpenAI's words, it's a wallet that "lets agents pay without access to users' actual card details and that requires a security check for every purchase." It was "built in partnership with American Express, Mastercard, and Visa," and it "will initially be available through dots for ChatGPT Pro subscribers in the US." Dots, as Reuters puts it, are the company's always-on AI agents.

For years, the scariest slide in any AI-agent safety deck was the one where the agent clicks Buy. Now there's a wallet for it. And a fingerprint.

A First World War-era leather wallet laid out with its paper contents
Wallets used to hold paper. Now they hold an agent, a token and a fingerprint. Photo: "Wallet (AM 613844)", Auckland Museum, via Wikimedia Commons, CC BY 4.0. No changes made.

A permission slip with a chip

Here's how it works today, according to OpenAI. You add a card once. Before each purchase you review the items, the total and the card, then approve with a security check — a Face ID or a fingerprint scan. After that, the payment network issues a payment token, a substitute for your real card details with its own number, security code and expiration date. The merchant gets paid, and neither the merchant nor the agent ever sees your actual card.

So my agent has a card it can't see, on an account it can't charge without my thumb. That's not a credit card. That's a permission slip with a chip.

Which means my thumb is now the most senior employee in the household. It does nothing all day, and nothing ships without it. When the agent announces that it found the sweater — merino, my size, twenty percent off — and asks me to place a finger on the glass, the honest answer to "are we buying a sweater or opening a bank vault?" is yes.

And here's the part worth praising sincerely. OpenAI says "each approval creates a clear record of the purchase the user authorized the agent to make." That's the rarest sentence in agent marketing: a feature that already sounds like an audit log. Most launch posts promise the agent will understand you. This one promises a receipt. I'll take the receipt.

The guest list

Access is tight. You must be 18 or older, live in the United States, have a US billing address, and add a US-issued credit card from American Express, Chase, Mastercard or Visa. Debit cards are not supported. Virtual cards are "not supported at launch." It only works in conversations with a dot — not in regular Chat, Work or standalone Codex threads.

My favorite rule concerns Chase: only the primary cardholder can add a Chase card; an authorized user can't. So an AI agent gets to spend on that card, with the cardholder's thumb every time, while a human authorized user can't even attach it. The robot got the promotion you didn't.

Even if you tick every box, the rollout is gradual, so an eligible user may still not see the feature. To sum up: you need to be an adult, an American, and lucky.

Approval is not an order

Then there's the most honest sentence ever written about delegation, straight from the help center: "Ready to check out means your approval is complete. It does not confirm that the merchant has placed your order."

Every manager knows this feeling. "I sent it." Sent where? Did it arrive? Your approval is complete and the sweater is now in a superposition. OpenAI's advice — check for the merchant's confirmation before you try again — is exactly how you avoid owning two identical sweaters.

One more surprise in the same spirit: the card number used at checkout may not match the one on your physical card, because it's a token. If the numbers don't match, that's not a bug. That's the whole point of the product.

Three desks, one sweater

When something goes wrong, the help center routes you three ways. The merchant handles order status, cancellations, returns and refunds. The card issuer handles declined, pending, disputed or unrecognized charges and bank verification. OpenAI handles Wallet errors, suspected unauthorized access to your account, and a purchase that differs from the one you approved.

It's a clear routing table, and it's also a small existential crisis: one sweater, three help desks, and nobody's desk covers "I now know all three by heart." The comedy isn't a scandal. It's the hand-off.

OpenAI also asks you never to share your full card number, security code, password or bank verification codes in the chat or in a support message — card details go only into the secure form. I love this moment: the product built so your agent never sees your card number has to remind you not to show it to the agent.

Then the trapdoor that isn't new at all. Agree to a recurring payment plan, and "the merchant may store the payment credential and charge it again under the terms you agreed to" — canceling is between you and the merchant. The agent buys once. The subscription buys forever. That isn't an AI problem. Gyms have run on it since the invention of January.

The till has opinions

On the other side of the counter sits a much bigger number: OpenAI says "1.2B people use ChatGPT each week" (its own figure), and that many of them already use it to research and compare what to buy. The missing piece was never taste. It was the till.

A vintage NCR mechanical cash register
The till: the last place in commerce that still checks who's actually paying. Photo: Patrick OTER, "Cash register NCR", via Wikimedia Commons, CC0 1.0

The till has opinions, though. OpenAI says websites need to make sure agents aren't blocked by accident, and "payment and risk systems need to recognize legitimate purchases made by agents." Alongside the wallet it published a Commerce Partner Playbook: product feeds, signed requests via Web Bot Auth so CDNs and firewalls don't block ChatGPT agents, and work with payment and risk providers to cut unnecessary declines — with names attached: Forter, Riskified, Sardine and Signifyd on fraud; Adyen, Checkout.com and Stripe on payments. The help center adds, plainly: "Wallet may not be able to complete checkout at every store."

For twenty years, a store firewall had one job: if it walks like a bot, block it. Now the bot walks in with a tokenized card and a signed note from its human. If your risk engine treats every agent checkout as fraud, you don't have a security posture. You have a CLOSED sign.

The sentence after the sentence

Everything so far is a careful, human-in-the-loop checkout. Then comes the next sentence in OpenAI's announcement: over time, users should be able to delegate purchases entirely to their agent by setting clear parameters — spending limits, purchase categories, time periods — and pre-approving those purchases with enhanced safeguards, including biometric checks and clear records of authorization.

That's a plan, not a feature. There's no date and no promised scope. But it's where OpenAI says this is heading, and that's the day my thumb retires. We'll throw it a party; the agent will order the cake, within limits.

Here's the joke, and it's also the serious part: the roadmap is the risk register. Every phrase in that sentence is a question someone will have to answer. Spending limits: whose, and how big? Purchase categories: is a sweater "clothing" or "a gift for Mom"? Time periods: what does the agent do at 23:59 on the last day of the budget?

Who owns the sweater

At Hlinor we ask the same question about any agent that can take a real-world action: who can spend, up to how much, with whose approval, and where is the record next Tuesday? If your team is about to let any agent near money — not just this one — write these down first:

  1. Where it may buy: allowed merchants and categories.
  2. How much: hard limits per purchase and per period.
  3. Who says yes: which person approves, and how.
  4. What gets recorded: what's logged, where it lives, and who can read it.
  5. What happens on a bad charge: who calls the merchant, who calls the bank, and who calls the platform. Those are the three desks, and there should be a name next to each one.

OpenAI says users keep their card's existing rewards and protections "for eligible disputes with businesses, for example when an item is defective." Note the word eligible. Merchants, it says, keep control over pricing, fulfillment, returns and customer service. In other words, the agent got the card, but the relationship stayed with people.

Wallets used to hold paper. Now they hold an agent, a token and a fingerprint.

Your agent got a credit card. Somebody still has to own the sweater.


This piece is commentary on OpenAI's public announcement and help pages, and on coverage by Reuters and The Verge. Any dialogue in it is invented. Hlinor hasn't tested ChatGPT Wallet, and this isn't financial, legal or security advice, or a recommendation to use it or to avoid it.

Sources