An AI agent can now spend your money, but only after your face or fingerprint says yes. OpenAI launched ChatGPT Wallet on November 13, and its design answers who approved a purchase. It leaves open who pays when an agent buys the wrong thing.
How a purchase works
OpenAI built the wallet with American Express, Chase, Mastercard and Visa. It rolls out first inside dots, the always-on personal agents OpenAI introduced on September 29, and only for ChatGPT Pro subscribers in the US who are 18 or older and hold an eligible US-issued card.
A user adds a card once. Before each purchase, the user reviews the items, the total and the selected card. The user then approves with a security check such as Face ID or a fingerprint scan, and each approval creates a record of what the user authorized.
After approval, the card network issues a payment token that stands in for the real card, with its own number, security code and expiry date. The merchant and the agent never see the actual card details, and the charge lands on the card the user picked. OpenAI says users keep their card rewards and their protections for eligible disputes with businesses, such as a defective item.
The launch is deliberately narrow. OpenAI says it wants to learn from real use and work with businesses, banks and payment networks before it expands access. Its announcement mentions no fees for shoppers or merchants.
Who pays when an agent gets it wrong
OpenAI’s announcement does not say who carries the loss when an agent makes a mistake. Banking Dive reports that card networks and tech companies are still working out which of them bears responsibility.
The protections OpenAI names cover disputes with businesses. A user who approves the wrong cart is in a different position. This suggests a mistaken purchase that the user approved will look like any other authorized purchase, and the approval record will back that up.
American Express has a standing promise on agent errors. In April, it announced Agent Purchase Protection, which covers charges caused by a registered AI agent’s error when the cardholder authorizes the agent and the agent sends authenticated purchase intent to Amex. Luke Gebb, Amex’s EVP and Head of Global Innovation, says Wallet is the company’s first Agentic Commerce Experiences integration. OpenAI’s announcement does not say whether Amex’s protection applies to Wallet purchases.
The sharper test comes later. OpenAI expects users to hand purchases fully to their agent within limits such as a spending cap, purchase categories or time periods, with biometric checks and authorization records as safeguards. Once the tap moves from each purchase to a one-time pre-approval, “the user approved it” stops being a clean answer. My take: that is when issuers, networks and AI companies will have to write down who pays.
What the approval step does to conversion
OpenAI’s announcement gives no figures on how many approvals become completed orders, and it does not say whether merchants can see purchases that shoppers decline to approve. The approval step belongs to OpenAI’s wallet, not to the merchant’s checkout page, so retailers cannot tune it the way they tune their own checkout.
The bigger risk sits with merchants. OpenAI warns that some sites may unintentionally block agent purchases, and that risk systems may flag legitimate agent purchases and decline or cancel them. Banking Dive reports that Amazon and United Airlines have blocked some agentic activity on their sites.
OpenAI published a Commerce Partner Playbook to address this. It tells merchants to configure their CDN or firewall to recognize signed requests using Web Bot Auth, upload product feeds with current prices and availability, and work with their payment and risk providers. OpenAI says it is working with fraud-prevention firms Forter, Riskified, Sardine and Signifyd, and with payment providers Adyen, Checkout.com and Stripe, so their systems recognize legitimate agent purchases.
Volume will lag the launch. Banking Dive reports that executives and analysts expect meaningful agentic commerce volume in 2027 at the earliest. BCG’s Inderpreet Batra expects far more of it that year as personal agents gain adoption.
My take: for now, a tap costs a retailer less than a decline. Payments and retail teams should put their effort into agent access and risk rules, not into the approval screen.
What to watch next
ChatGPT Wallet is a controlled pilot limited to US Pro subscribers, but it sets a template other agent payment products may follow: a tokenized card, a human approval and a record of both. The open question is what happens to liability when the human steps back. Merchants and issuers that settle their answers before delegation arrives will be better placed when volume does.
The post ChatGPT Wallet: Card Networks Give AI Agents a Way to Pay, With a Biometric Gate appeared first on DataFLOQ.
