On 13 November OpenAI launched ChatGPT Wallet, which lets its dots agents buy things for a user. CNBC reported that it was built with American Express, Mastercard and Visa, and that the agent never sees the card: the user enters it once, approves each purchase with a security check, and the network issues a token tied to that one transaction. The Paypers reported the details of the first release: US ChatGPT Pro subscribers aged 18 and over with eligible US-issued cards, a biometric check on each purchase, and a record of what was authorised. Cardholders keep their rewards and their protections for eligible disputes, and OpenAI told CNBC it is not taking a cut of transactions to start.
The agent can be built by anyone. Whether it can spend is still decided by the bank whose name is on the card.
The same day, PYMNTS reported a Wall Street Journal story that JPMorgan Chase is stopping AI agents from shopping with its customers' credit cards when the customer hands the agent the card number, while cards from Bank of America, Capital One and Citigroup are already being used by agents when customers allow it. According to the report, JPMorgan has been slow to adopt the networks' agent tokens, and Jamie Dimon said: "You don't want rogue agents, and there could be thousands of them."
Put those side by side and there are now three ways to put a card in an agent's hand, and each is treated differently. The first is the raw card number pasted into an agent, which JPMorgan Chase has just refused, and which several banks, per the same report, plan to treat like a card lent to a family member: the charges belong to the cardholder. The second is a single-use virtual card, the route Stripe's Link uses; it works, but merchants told the Journal it is classed as a commercial card, costs more to accept and makes returns harder. The third is a network token bound to the agent and the purchase, which is what OpenAI's wallet uses and what Visa described in June as tokens bound to specific agents and use cases, inside spending caps and approval thresholds the customer sets.
Only the third gives the issuer what it wants: a credential it can see is an agent's, limited to what the customer approved, with an authorisation record behind it. That is why the launch partners are the networks, and why American Express, through an executive quoted by CNBC, described the rollout as starting in a contained environment.
This settles something we have been circling for two months. In October we asked who carries the loss when an agent buys the wrong thing, and then noted that the first business deployments answered by letting the agent execute only what a person had approved. The consumer side has now landed in the same place. Every purchase is approved, recorded and tokenised, and the party that decides whether the payment goes through is still the bank that issued the card.
The Paypers reports that OpenAI expects users eventually to delegate purchases within limits, categories or time periods they set in advance. When that happens the approval record becomes the thing every dispute is settled against, and the issuer's view of that record becomes the gate on how far delegation can go.
Sources
- 01CNBC: OpenAI launches ChatGPT Wallet, allowing its dots AI agents to make purchases for users (2026-11-13)
- 02The Paypers: OpenAI launches ChatGPT Wallet for agent payments (2026-11-16)
- 03PYMNTS: JPMorgan stops AI agents from using customer credit cards (2026-11-13)
- 04Visa: How Visa is partnering with OpenAI to build the future of agentic commerce (2026-06-10)
Tell us what you are building. Every note gets a real read.