The payment step is the tidy part
The design is careful about the obvious fear. The agent does not receive the user's actual card details. The user adds a card once, the agent assembles a purchase, and then the process stops. The user sees the items, the total and which card will be charged, and approves that specific purchase with a security check such as Face ID or a fingerprint scan. Only after that approval does the network issue a token that stands in for the card at checkout.
That structure answers two questions at once. The agent never holds the user's actual credentials, so a compromised or badly behaved agent has less to leak. And every approval creates a clear record of the purchase the user authorized, which is the evidence a user would want if a charge is later disputed. OpenAI is explicit that existing card rewards and dispute protections for eligible disputes still apply.
Read the limits of that design, though. Approval is per purchase, today. The version where a dot buys on its own within limits you set in advance does not exist yet. OpenAI describes full delegation as a future capability: users setting parameters such as spending limits, purchase categories or time periods, pre-approving those purchases, with enhanced safeguards. What launched is an agent that shops and a human who still approves at the till.
OpenAI's own post names the real obstacle
The most useful paragraph in the announcement is not about tokens. It is the one telling businesses to make sure their sites do not block the agent.
OpenAI says websites need to ensure agents are not inadvertently blocked, and that payment and risk systems need to recognize legitimate agent purchases. It concedes some sites may unintentionally block Wallet purchases, and that risk systems may flag them, causing declines or cancellations. Alongside Wallet it published a Commerce Partner Playbook telling businesses to claim their listing, upload a product feed, configure firewalls to recognize signed agent requests, and talk to their risk providers.
That is a long list of things other companies have to do before the feature works smoothly. TechCrunch adds the sharper version: some retailers block AI agents deliberately, and it identifies Amazon as the largest retailer to ban them, in favor of its own agent. Whether a block is accidental or strategic, the result for the user is the same. The wallet opens, and the shop door stays shut.
What to test before trusting it
There is no merchant list in OpenAI's announcement, no count of how many Pro subscribers get access in the first wave, and no date for other plans or countries. OpenAI warns that blocking and risk declines can occur, but publishes no acceptance or failure rate. Those omissions matter more than the token design: a payment method is only as useful as the checkouts that accept the buyer presenting it, and here the buyer is software.
So the practical test is narrow. If you are in the first wave, try Wallet on an ordinary, low-value purchase at a retailer you already use, and watch where it fails: at agent access, at the risk check, or not at all. A successful purchase establishes that one transaction completed under those conditions; it does not establish consistent acceptance at that store or anywhere else.
Sources
- [1] OpenAI, "Introducing ChatGPT Wallet" (November 13, 2026)Read source
- [2] OpenAI Developers, Commerce Partner PlaybookRead source
- [3] TechCrunch, "ChatGPT's new Wallet lets AI agents shop and check out for you" (November 13, 2026)Read source