United Airlines spent November 11 explaining why it will block personal AI agents from booking its flights. Its release warned that “actions taken by these tools may result in unintended purchases, account changes, or use of MileagePlus miles.” Two days later, OpenAI gave its agents a way to pay.
ChatGPT Wallet launched on November 13 for U.S. Pro subscribers over 18. It lets OpenAI’s dots agents buy things on a user’s behalf. You enter a card once, ask your dot to find something, and approve each purchase with Face ID or a fingerprint. OpenAI announced it as a product built with American Express, Mastercard and Visa, and CNBC reports it is not taking a cut of transactions to start.
The biometric prompt is the whole safety design at launch. So the useful question is what that prompt covers, and what it quietly hands to you.
What the agent never sees, and what you sign off on
The card protection is real. After you approve, the network issues what CNBC describes as “a scoped card number, security code and expiration date that is specific to a given transaction.” Neither the dot nor the merchant sees your actual card number. The Next Web adds that OpenAI is working with Adyen, Checkout.com and Stripe on processing and with Forter, Riskified, Sardine and Signifyd on fraud.
Before payment, Wallet shows the merchant, the items, the total and the card you picked. The Neuron’s reading of OpenAI’s Wallet terms says the approval covers only those details, and if a material detail changes before the order goes through, Wallet has to ask again. That is a good rule. It stops a dot from swapping a cheaper flight for a pricier one after you looked away.
Here is what the screen does not show: whether the item does what you asked for. Face ID confirms that you, the account holder, accepted this basket at this price. It says nothing about whether the 65-inch TV in the basket works with your kid’s console, or whether the hotel has the free cancellation you mentioned three messages earlier.
That gap is where agents fail. In Visa’s shopping benchmark earlier this month, the agents that failed still made a purchase in 54% of those runs, buying something that missed the brief or buying before they had found all the requirements. Wallet’s approval step catches the second kind only if you read the basket against your own list. It does not catch the first kind at all.
Make the approval screen do the checking
The approval moment is the only control you get, so make it carry the requirements, not just the price:
- Give the dot your must-haves as a list, up front. Budget ceiling, compatibility, delivery date, return window. In Visa’s test, the five models that uncovered the most hidden requirements passed 70.2% of runs against 44.4% for the five that uncovered the fewest, and you can skip the drawing out by handing them over.
- Ask for a confirmation line before it reaches checkout. Something like: “Before you request payment, list each of my requirements and the product detail that meets it.” If a line comes back blank or vague, do not approve.
- Read the merchant name, not just the total. A dot that cannot buy from the retailer you expected may find a reseller. That is a different return policy and a different customer service desk.
- Treat a second approval prompt as a signal. If Wallet asks again for the same order, something material changed. Find out what before you touch the sensor.
- Wait for the merchant’s confirmation. The Neuron notes that approval does not guarantee the order: the issuer still has to authorize it and the merchant still has to accept it, and some retail fraud systems reject agent purchases. A Face ID success is not a receipt.
Where a bad purchase goes
This is the part most early coverage skipped. Per The Neuron’s summary of the terms, OpenAI routes problems three ways, and the route decides how much protection you have:
| What went wrong | Who you contact | What backs you |
|---|---|---|
| Wrong item, or item never arrived | The merchant | The merchant’s return policy; for non-delivery, the card dispute process |
| A charge you did not authorize, or a billing error | Your card issuer | Existing card dispute rights |
| Wallet malfunctioned, or the charge differs from what you approved | OpenAI Support | No promise of automatic reimbursement |
Routing per The Neuron’s summary of OpenAI’s Wallet terms; the “what backs you” column is this site’s reading.
The table hides a trap. Once you approve with your face, “unauthorized” is a hard claim to make. According to The Neuron’s summary of the Wallet privacy notice, OpenAI logs the request, what you approved or rejected, the authentication result and what went to the merchant. That record protects you if a charge differs from the basket. It works against you if you approved the wrong TV.
A wrong but delivered item that you approved is usually a return, not a billing error. The CFPB says being charged when you did not receive what you ordered can be a billing error, but a dispute over the quality of what arrived follows a different process with conditions attached: a good-faith attempt to settle it with the seller, a price over $50, and a purchase in your home state or within 100 miles of home. When you do have a billing error, the clock matters: you must send a written notice within 60 calendar days after the charge appeared. A phone call to the issuer does not start that protection.
OpenAI’s broader Service Terms, again per The Neuron, make users responsible for the transactions their dots make. Section 4 of the Wallet terms says it plainly: “AI systems can make mistakes.” Plan on the mistake being yours to unwind.
The cards it will not reach, and the companies that should care
Wallet works only where merchants let agents in, and many do not. Besides United, CNBC reports that Amazon prevents agents from purchasing on its platform, and OpenAI has published guidance so merchants avoid “inadvertently blocking OpenAI’s agents.” The opposite camp is building permission layers: Sierra’s Poppy 1.0 protocol is now on every Shopify store, and Meta’s Muse checks for it first. We covered that access fight in Amazon blocking shopping agents while Meta built a standard to let them in.
For employers, the issue is narrower and more immediate. Wallet sits on a personal Pro subscription, and no launch coverage mentions admin controls, spending policy or audit export. If an employee adds a corporate card to their own dot, the company inherits a purchasing channel its expense policy never contemplated. Coverage does not say whether corporate cards are eligible, so check before you assume they are not, and write the rule now. Our earlier look at when to connect dots to work accounts applies here with a card attached.
When the Face ID prompt goes away
Per-purchase approval is a launch setting, not the destination. The Neuron reports OpenAI wants users to delegate purchases through “predefined spending limits, permitted categories, or time periods,” and TNW says OpenAI expects users to hand purchases to their agent entirely over time. Amex’s Luke Gebb told CNBC it makes sense to “start it in a contained environment, and then keep rolling from there.”
That is the moment Visa’s 54% figure stops being a benchmark result and starts arriving on statements. When delegation ships, set the narrowest version first: a low per-purchase cap, one category you can return easily, and a short time window. Keep anything with a nonrefundable fare or a restocking fee on manual approval. The agent payment stack we mapped in August already offers spending limits at the card and wallet layer; use those before trusting an instruction in a prompt.
Until then, the Face ID tap is the only check between your dot’s judgment and your card. It is worth the five seconds it takes to read what you are agreeing to.
