On 10 June, at its Payments Forum, Visa wired its network into ChatGPT. In the AP's words, it "embedded its payment network inside of ChatGPT" so an AI agent can find things and buy them for you. [E1] The demo is deliberately boring: ask for wireless headphones under $150, and the agent finds a pair and buys it, inside limits you set — spending caps, allowed merchants, a prompt to approve. [E2] In principle it reaches any of the 175 million-plus places that take Visa.
Underneath is new plumbing for machines to pay. Visa's Trusted Agent Protocol signs each agent cryptographically, so a shop can tell an approved shopping bot from a scraper or a fraudster. [E3] That's the real move. It's not that a chatbot can shop. It's that the card network is building an ID system for software that spends money.
Which raises the question in the headline: who's on the hook when the agent gets it wrong? Old way — you clicked, the store saw you, the network approved the card. New way — ChatGPT picks the thing, Visa checks the agent and scores the risk, the store ships it. McKinsey says the quiet part: the "customer" is now an AI agent, and it needs permission you can prove. [E4] So fraud stops being "did the cardholder click?" and becomes "what exactly did you let the agent do?" Nobody has written that contract yet.
This is our own premise going commercial — agents that act, not just answer — and the interesting asset is the tollbooth. Whoever owns the agent-ID layer taxes the whole thing, which is why Mastercard shipped its own "Agent Pay for machines" the same month. [E5] One catch worth keeping in view: Visa makes a human approve each purchase for now, and OpenAI's last "buy in ChatGPT" button charged merchants 4% and got quietly killed in March for lack of use. [E1] The rail is real. Whether anyone lets the robot press "buy" is the open question.