The plumbing, and the question it opens
When OpenAI unveiled ChatGPT Wallet in November 2026, the framing was about a narrow technical problem: how to let an agent make a purchase without giving the agent, or the merchant, access to the customer's actual card number. OpenAI described it as the first wallet that lets agents pay without access to users' actual card details and that requires a security check for every purchase, while preserving existing card rewards and protections for eligible disputes. The build partners are American Express, Mastercard, and Visa, and the rollout starts through dots, OpenAI's agent platform, for ChatGPT Pro subscribers aged 18 and older in the US with eligible US-issued cards. The weekly active audience OpenAI cites is 1.2 billion people using ChatGPT.
The mechanics run like this. A user adds a card once, then approves each purchase without re-entering payment details. Before the charge, the user reviews the items, the total, and the selected card, then confirms with a security check such as Face ID or a fingerprint scan. Each approval produces a clear record of the purchase the user authorized. After approval, the payment network supplies a payment token that stands in for the card details, and the card is charged without those details ever reaching the business or the agent.
Who owns the dispute
The important line for an operator is that OpenAI deliberately did not redesign the buyer-seller relationship to run through the agent. Users transact directly with businesses, and the merchant keeps its customer relationship, plus control over pricing, fulfillment, returns, and customer service. So when an item arrives defective, the dispute sits between the customer and the merchant, not between either of them and OpenAI or the agent that placed the order.
The customer retains the card protections they would normally have. OpenAI says users keep their card's existing rewards and protections for eligible disputes with businesses, for example when an item is defective. The banks and networks bring their own expertise to evaluating transactions, which is the same framing OpenAI uses for how a cardholder's claim over a flawed order gets assessed. In practice, that means the evaluation of such a claim depends on banks and networks bringing their own expertise to the process.
That is not a minor distinction. It means the merchant, not the agent, bears the obligation to deliver what was sold or make it right. An agent is the conduit, and the token is the payment method. Neither the agent nor OpenAI has stepped into the shoes of the party that was supposed to ship the goods. The merchant therefore is the party a defective-order dispute is filed against, and the merchant keeps its leverage over refunds, replacements, and service.
Who carries the fraud liability
Fraud exposure in this model splits into two directions, and the sources make the division explicit. On one side, OpenAI is building the system to reduce the chance that legitimate agent purchases get flagged. Its Commerce Partner Playbook tells businesses to work with payment and risk providers so a real agent-led order is not declined or cancelled. The company is working with has not announced a blanket liability arrangement under which it would automatically compensate users for every problem arising from an agent-led purchase, and that applicable card protections and the circumstances of the dispute will remain relevant. So the fraud-loss question, as it stands, goes back to the card-network dispute framework. An inference is that liability is settled under those card rules rather than by who pressed buy, though the sources describe the process without spelling out the exact split.
The practical upshot for an operator
The cleanest way to hold the answer is to separate the three layers. The payment is tokenized and authorized by the user. The commercial relationship and goods obligation stay with the merchant. The evaluation of disputes and fraud runs on the card networks' existing rules. An operator should treat an agent-placed order as a normal card transaction for these purposes, with one added hazard. Because risk systems may still treat an agent's checkout as suspicious, a defective-order dispute can get tangled with a decline or a fraud flag, even when the customer properly authorized the purchase.
So the answer to the title comes out this way. The defective-item dispute belongs to the customer against the merchant, pursued through the cardholder's normal card protections, and it is not a problem OpenAI agrees to automatically cover. The company has not announced a blanket liability arrangement under which it would automatically compensate users for every problem arising from an agent-led purchase, and the applicable card protections and the circumstances of the dispute will remain relevant. The agent changed who assembles and submits the order. It did not change whose obligation it is to ship the right product.
Sources
- Introducing ChatGPT Wallet (OpenAI, 2026-11-13)
- ChatGPT Wallet: How AI agents can shop and pay for you and who is liable if something goes wrong (CNBC TV18, 2026-11-15)