OpenAI is wiring a payment network into ChatGPT. Under an expanded partnership announced at the Visa Payments Forum on Wednesday, AI agents inside OpenAI’s products will be able to shop and pay on a user’s behalf at, in principle, any of the more than 175 million merchant locations that accept Visa, once the user grants permission.
The pitch is simple: tell ChatGPT to find wireless headphones under $150 or reorder paper towels, and the agent completes the purchase. Visa supplies the plumbing, tokenised card credentials bound to a specific agent, real-time authorisation, agent identification, and fraud monitoring, the same machinery it runs across more than 300 billion transactions a year.
The user sets the limits: spending caps, approval thresholds, and merchant restrictions, so a human stays in command, at least at first.
OpenAI’s second swing at commerce
This is not OpenAI’s first attempt to turn ChatGPT into a checkout. Its earlier Instant Checkout, launched late last year, let the chatbot find and buy a specific item, but it leaned on a 4 per cent merchant fee that retailers balked at, saw little adoption, and was retired in March.
The Visa deal offloads the part OpenAI struggled with, the trust, fraud, and dispute machinery, to a network built for exactly that. “Making the leap from having AI agents recommend what to buy to doing the purchasing requires a whole different level of trust,” Visa’s chief product and strategy officer Jack Forestell said.
It is also a land grab. Visa already runs a Trusted Agent Protocol with Microsoft, Stripe, Shopify, and Worldpay; Mastercard has its rival Agent Pay; Google has Universal Cart; and Amazon is selling its own shopping AI. Even brokerages are letting agents spend. Everyone wants to own the moment an AI hits “buy”.
For now, though, it is more promise than product.
Visa and OpenAI disclosed no launch date, no pricing, and no user interface, and gave no detail on what merchants or customers will pay. Visa’s own product page spells it out: the system is “currently in the process of deployment” and the final version “may not contain all of the features described”.
The unresolved questions are the hard ones: who eats the loss when an agent buys the wrong thing or a user disputes a charge, how banks treat fraud claims on agent-initiated payments, and whether people will really let software check out without looking.
Forestell framed the end state himself: you approve a thousand agent purchases, and then “your agent says, ‘Do you want me to just not check?’” Whether shoppers say yes is the whole bet.
