ChatGPT Wallet lets an agent pay without seeing your card. The design choice is the approval on every purchase
OpenAI built ChatGPT Wallet with American Express, Chase, Mastercard and Visa. The agent gets a network token, the user approves each purchase biometrically, and merchants must learn to accept agents.

Picture the first purchase an agent makes for you. It found the replacement part, compared three sellers and filled the cart. The question that matters is not whether it can pay. It is what exactly you said yes to.
OpenAI’s answer, in ChatGPT Wallet, announced on November 13, is narrow on purpose: you say yes to one purchase at a time.
How a purchase works
ChatGPT Wallet was built with American Express, Chase, Mastercard and Visa. It is rolling out inside dots, OpenAI’s agent experience, initially for ChatGPT Pro subscribers in the US who are 18 or older and have eligible US-issued cards.
The user adds a card once. Before each purchase, they see the items, the total and the selected card, and approve with a security check such as Face ID or a fingerprint. Only then does the payment network issue a payment token, a substitute card number, security code and expiration date, which is charged to the real card. Neither the agent nor the merchant sees the actual card details. Each approval creates a record of what the user authorized.
- Agent builds the cart
- User reviews and approves
- Network issues a token
- Merchant charges the token
Users keep their card’s rewards and protections for eligible disputes. Merchants keep the customer relationship, pricing, fulfillment, returns and service. Card issuers and networks keep evaluating each transaction as they do today.
The authority sits with the purchase, not with the agent
Most debates about agentic commerce assume the hard problem is giving an agent a payment method. ChatGPT Wallet shows a different design: the agent never holds a payment method at all. It holds the ability to ask. The authority to spend is created at the moment of approval, scoped to one basket, and expressed as a token the network controls.
That is the pattern we described as reasoning that travels while the authority to commit stays put: the model can plan, compare and prepare, but the commitment is bound to an explicit act by the principal and enforced by infrastructure outside the model.
OpenAI says it expects, over time, to let users delegate purchases entirely, with parameters such as spending limits, categories or time periods, and pre-approval with enhanced safeguards. That is a plan, not a feature. When it arrives, the per-purchase yes becomes a standing mandate, and the questions shift: who set the limit, how it is revoked, and what record shows the agent stayed inside it. Those are the questions any enterprise will face the day a procurement agent is allowed to buy without a click.
The merchant side is the harder half
The less discussed part of the announcement is addressed to businesses. OpenAI admits that some sites “may unintentionally block” purchases through ChatGPT Wallet, and published a Commerce Partner Playbook with four kinds of work:
- Product data: claim the business in ChatGPT, then keep a product feed with current prices and availability.
- Agent access: configure the CDN or firewall to recognize signed agent requests using Web Bot Auth, so bot defenses do not turn away legitimate agents.
- Risk systems: payment tokens work with existing checkouts, but fraud models may decline legitimate agent purchases. OpenAI names fraud partners Forter, Riskified, Sardine and Signifyd, and payment providers Adyen, Checkout.com and Stripe, as working on recognizing these transactions.
- Agent-ready actions: plugins for availability and carts, or WebMCP tools on the site.
For a retailer, a bank or an airline, this is the same tension we noted when companies started turning agents away at the front door: the controls built to stop automated abuse now also stop customers who send an agent. Telling a good agent from a bad bot becomes an identity problem, and signed requests are the start of an answer, not the whole answer.
What enterprises should take from a consumer launch
ChatGPT Wallet is a US consumer product in limited availability, and OpenAI does not state any fees in its announcement. But the architecture is a reference for any company letting agents transact:
Separate preparing from committing. Let the agent assemble the transaction; bind the commitment to a person, a policy or a system that is not the model.
Keep credentials out of the agent. A token issued per transaction is less valuable to steal than a stored card or an API key with spending rights.
Make every approval a record. When something goes wrong, the question will be what was authorized, by whom and when. Design the answer before the first dispute.
Check your own front door. If your company sells online, find out today whether your bot defenses, fraud rules and checkout would accept a legitimate agent purchase, and whether you can tell it apart from abuse.