OpenAI Gives Its AI Agents a Wallet: ChatGPT Can Now Pay for Your Shopping

OpenAI has given its artificial intelligence agents a wallet. On Friday 13 November, the company launched ChatGPT Wallet, a payment feature that lets its always-on agents, known as dots, complete purchases on a user’s behalf without ever seeing the user’s actual card number. The launch, built in partnership with American Express, Mastercard and Visa, is the clearest sign yet that the world’s biggest AI company wants to move from helping people choose what to buy to actually buying it for them.

The wallet arrives at a moment when the race to control agent-led shopping is accelerating across the whole technology industry. Google, Microsoft, Meta and Perplexity already let their assistants complete purchases inside a conversation. OpenAI, whose chatbot is used by 1.2 billion people each week according to the company, had until now been the odd one out: it handed checkout back to merchants in March after its first attempt at in-chat payments failed to win them over. ChatGPT Wallet is its second attempt, and this time it is built on the card networks’ own rails rather than on top of them.

For now the feature is deliberately small. It is rolling out only to ChatGPT Pro subscribers in the United States who are aged 18 or older and hold eligible US-issued cards, and only inside dots conversations. But the ambition behind it is large. OpenAI says it expects users eventually to hand purchasing over to their agents entirely, setting spending limits, purchase categories or time periods in advance. The question is whether the shops, the banks and the fraud systems that actually process payments will let the agents in.

How a purchase works

According to OpenAI, ChatGPT Wallet is “the first wallet that lets agents pay without access to users’ actual card details and requires a security check for every purchase”. The claim is the heart of the product. A user adds an eligible payment card once. When a dot is asked to buy something, it finds the product, and before any money moves the user is shown the items, the total and the selected card. The user then approves the purchase with a security check such as a Face ID scan or a fingerprint.

Only after that approval does the payment network step in. The network supplies a payment token, a substitute for the user’s real card details, carrying its own card number, security code and expiry date. The selected card is charged, but neither the business nor the agent ever sees the actual card number. Each approval also creates a clear record of the purchase the user authorised the agent to make, according to OpenAI.

The design is deliberately conservative in one important respect: at launch, no purchase can happen without the user looking at it first. OpenAI says it expects that to change over time, with users able to delegate purchases entirely to their agent by setting clear parameters, such as spending limits, purchase categories or time periods, and pre-approving those purchases with enhanced safeguards that build on the biometric checks and authorisation records already in place. No date has been given for that step.

Users keep what they already had. OpenAI says cardholders retain their existing card rewards and the protections that apply to eligible disputes with businesses, for example when an item arrives defective. The businesses, meanwhile, keep their customer relationships and control over pricing, fulfilment, returns and customer service. The user transacts directly with the business, and the agent is a go-between rather than the seller. That distinction matters, because it is the opposite of the approach OpenAI tried first.

Who gets it first — and who is left out

ChatGPT Wallet is starting with a narrow group. According to Reuters, which reported the launch on Friday, it will initially be available in the US to ChatGPT Pro subscribers aged 18 and older with eligible US-issued cards, and it can be accessed through dots, the company’s always-on AI agents. OpenAI’s own announcement says the same, adding that the rollout will remain limited while the company learns from real-world use and works with businesses, banks and payment networks to improve compatibility and reliability.

The cards at launch are those issued by American Express, Mastercard and Visa, the three networks that built the wallet with OpenAI. The Verge, reporting on OpenAI’s help pages, also lists Chase cards as eligible, and OpenAI’s help centre names American Express, Chase, Mastercard and Visa. Chase cardholders must be the primary account holder, according to the same help material. Users must also have a US billing address, and the feature works only inside dots conversations, not in ordinary ChatGPT chats, Work conversations or standalone Codex threads.

That narrowness is not an accident. Dots themselves are only about six weeks old. OpenAI introduced them on 29 September 2026 as personal agents that keep working in the background when a conversation is closed. Each dot runs on its own cloud computer, powered by the GPT-6 Astra model, and can browse websites, read pages, open files and use apps without touching the user’s own device. Dots connect to more than 4,000 apps through the ChatGPT plugin ecosystem, according to OpenAI. They are a feature of ChatGPT Pro and Business Premium plans, with Pro listed from $100 a month, and they are not available in the European Economic Area, Switzerland or the United Kingdom at launch.

The limited rollout also reflects how much of the payments system still has to be taught to recognise an agent. OpenAI says some websites may unintentionally block purchases made through ChatGPT Wallet, and that risk systems may flag legitimate agent purchases as fraud, leading to declines or cancellations. The company is, in its own words, “starting with limited availability to ChatGPT Pro users, so we can learn from real-world use and work with businesses, banks, and payment networks to improve compatibility and reliability”.

The plumbing behind the wallet

What makes ChatGPT Wallet different from a chatbot with a saved card is that it sits on infrastructure the card networks themselves built for agents. Visa launched its Intelligent Commerce programme in April 2025, and Mastercard launched Agent Pay in the same month. Both were designed for a world in which software, not people, presents payment credentials at checkout. OpenAI’s wallet uses the networks’ tokenisation: the substitute card number is issued by the network, the bank or network still makes the authorisation decision, and the merchant’s existing checkout form receives a token it already knows how to process.

OpenAI has also published a Commerce Partner Playbook alongside the launch, with tools and practical guidance to help businesses support customers who shop through agents. The playbook covers making product information available, ensuring agent access and working with payment and risk providers to reduce unnecessary declines. Businesses can claim their business in ChatGPT and, once verified, update their details and upload a product feed with current pricing and availability. They can configure their content delivery network or firewall to recognise signed requests from ChatGPT’s cloud browser using Web Bot Auth, and confirm that their general security measures are not inadvertently blocking the agents. They can build plugins so agents can check availability, manage carts or take other actions through their existing systems.

Behind the scenes, OpenAI says it is working with fraud prevention companies Forter, Riskified, Sardine and Signifyd, and with payment service providers Adyen, Checkout.com and Stripe, to help their systems recognise legitimate purchases made by agents. The point of that work is to stop a real purchase from looking like a stolen card. An agent buying from a cloud browser, at machine speed, with a tokenised card, looks to a fraud system very much like the behaviour fraud systems were built to stop. OpenAI’s own announcement warns businesses that risk systems “may still flag legitimate purchases made by agents, leading to declines or cancellations”.

There is also a commercial detail that signals how OpenAI is positioning itself. According to CNBC, OpenAI is not taking a percentage cut of transactions at launch, and is focusing instead on establishing relationships with online sellers. That is a deliberate contrast with the commission-based model of its first attempt. American Express, for its part, is moving cautiously: Luke Gebb, the company’s executive vice president and global head of innovation, told CNBC that Amex is testing the feature in a contained environment before expanding, according to the Silicon Report’s account of the interview.

OpenAI’s second attempt at the checkout

ChatGPT Wallet is not OpenAI’s first try at putting a payment step inside ChatGPT. It is the second, and the differences between the two attempts explain a great deal about what the company has learned.

The first attempt was Instant Checkout, launched on 29 September 2025 with Stripe, starting with US sellers on Etsy. It was built on the Agentic Commerce Protocol, an open standard for letting AI agents buy things, and it kept the purchase inside the chat: the user never left ChatGPT. Within months, OpenAI changed course. On 24 March 2026 it wrote that the first version “did not offer the level of flexibility” merchants needed, and said it was allowing merchants to use their own checkout experiences while it focused its efforts on product discovery. The in-chat checkout was retired, and ChatGPT became a place to find and compare products rather than a place to pay for them.

By October 2026, four rival assistants could complete a purchase inside the conversation — Google’s AI Mode and Gemini, Microsoft Copilot, Meta’s Muse and Perplexity — while ChatGPT, the assistant most people associate with AI shopping, handed checkout back to merchants, according to an agentic commerce tracker published that month. OpenAI’s shopping experience had been rebuilt around discovery: product results, business profiles and merchant checkouts in a new tab or an in-app browser.

ChatGPT Wallet reverses that retreat without repeating the original mistake. The purchase still finishes in the merchant’s own checkout, using the merchant’s existing payment form. The merchant stays the merchant of record. The card keeps its rewards and dispute rights. The issuer makes the authorisation decision. The token goes into the merchant’s existing checkout form. As one payments industry analysis put it, OpenAI has stopped trying to replace the payments plumbing and is now building on the networks’ plumbing. The company itself says it is “building agent-led commerce to fit reliably into existing systems”.

The timing is also shaped by the calendar. The launch landed two weeks before Black Friday on 27 November, the biggest shopping event of the year in the United States, when the merchants, fraud teams and payment networks that ChatGPT Wallet depends on will be under their heaviest load. A limited rollout to Pro subscribers means the first real-world transactions will be few, but they will arrive in the most demanding possible week for the systems that have to recognise them.

The shops that are already saying no

The largest obstacle in front of ChatGPT Wallet is not technical. It is that some of the biggest sellers on the internet do not want AI agents buying things on their platforms at all, and have already moved to stop them.

Amazon, the largest retailer to ban AI agents, prevents them from making purchases on its platform, according to TechCrunch. The company would rather shoppers use its own Alexa agent than a rival’s. United Airlines said on Wednesday 11 November that it will block agents from booking tickets, according to CNBC, and Delta bans them too, according to Engadget. JPMorgan and Charles Schwab block agents from moving customer money, according to CNBC and Mashable. The pattern is consistent: the companies that own the customer relationship are not willing to hand it to someone else’s agent.

Not everyone is blocking. Walmart allows agents to shop through agent profiles with their own spending limits, according to CNBC. That is the model OpenAI is hoping spreads: agents with their own identities, their own limits and their own audit trails, welcomed as a new kind of customer rather than treated as a bot to be filtered out.

There is also a quieter problem that no playbook can fix on its own. AI agents are still unreliable shoppers. Visa tested 22 AI models as shoppers this month and they failed 42.8 per cent of the time, according to The Next Web. A failed purchase in this context does not just mean a declined card. It can mean the wrong size, the wrong colour, the wrong quantity, or a purchase from a seller the user would never have chosen. Every one of those failures lands on the merchant’s customer service team, and on the user’s patience. That is why OpenAI’s insistence on a human review and a biometric check before every purchase is not just a safety feature. It is the only thing standing between the wallet and a wave of disputed orders.

Privacy is the other open question. OpenAI’s privacy notice for the wallet says the company collects detailed purchase records, including merchants, items, prices and shipping information, and that it will not sell the data or share it for targeted advertising, according to reports on the notice. The company will, however, hold a record of what its users buy, through which merchants, and at what prices. For a company whose chatbot already knows what 1.2 billion people ask about each week, that purchase history is a new and commercially valuable layer of knowledge about its users.

A race the whole industry is now running

ChatGPT Wallet does not arrive in an empty field. The idea that software agents will soon do a meaningful share of the world’s shopping is now the working assumption of every large technology and payments company, and each of them is building its own version of the rails.

Visa’s Intelligent Commerce and Mastercard’s Agent Pay, both launched in April 2025, give agents a way to pay with tokenised credentials and spending controls set by the cardholder. Google’s Gemini cannot make transactions with Google Wallet, according to Google’s own help pages, but Google’s AI Mode can complete purchases in conversation. Microsoft’s Copilot can check out. Meta’s Muse, the always-on personal agent Meta launched on 8 September 2026, is the product OpenAI’s dots were most directly answering when they arrived three weeks later. Perplexity launched its own always-on agents, called Teammates, on 5 November 2026. Anthropic open-sourced a Claude Commerce Agents blueprint in September 2026, formalising the structure in which the merchant stays the merchant of record and the agent works through the merchant’s own systems.

The shared direction is unmistakable, but the standards are not settled. Google, Microsoft, OpenAI and Anthropic have also backed a shared signal in robots.txt that lets website owners express a preference against AI training, a separate effort that shows the four companies can agree on a common mechanism when it suits them. No equivalent shared mechanism yet exists for agent payments. Each network and each AI company is building its own version, and merchants are being asked to recognise all of them.

That fragmentation is the real contest. Whoever controls the moment of payment controls the customer relationship, the data about what was bought, and eventually the fee for making it happen. OpenAI is not taking a cut now, but the history of every wallet, marketplace and app store suggests that a cut follows once the volume does. The networks know this too, which is why Visa and Mastercard moved first, in April 2025, to make sure the agent economy runs on their rails rather than on a technology company’s.

Analysis: Why It Matters

The wallet is the moment the agent stops being an adviser

Until now, the most widely used AI assistant in the world could recommend but not buy. That distinction mattered more than it seemed. A recommendation is free advice; a purchase is a commitment of money, and whoever makes it takes on the risk, the liability and the relationship with the seller. By giving dots a way to pay, OpenAI has crossed the line from an information product to a transaction product. The company is no longer just shaping what people want. It is building the mechanism by which they get it, and that mechanism carries a record of every purchase it touches. The commercial value of that record, and the strategic value of sitting between the shopper and the shop, is the real prize here, not the feature itself.

The token is a genuine piece of engineering, not a marketing claim

It is easy to be cynical about a launch that amounts to a saved card with extra steps, but the token design solves a real problem. Handing an AI agent a credit card number would be reckless: the number could be logged, leaked, replayed or read by the model itself. A network-issued token with its own number, security code and expiry date, valid for one transaction and useless afterwards, means the agent never holds anything worth stealing. Combined with a biometric check on every purchase, it is a credible answer to the first question any sensible user would ask, which is what happens if the agent goes wrong. The design also matters for the banks, because it leaves the authorisation decision with the issuer and the network, where the fraud expertise already sits. OpenAI has not tried to become a payments company. It has asked the payments companies to do what they already do, for a new kind of customer.

The merchants’ veto is the story, not a footnote

The most important fact about ChatGPT Wallet may be the list of companies that will not accept it. Amazon, United Airlines, Delta, JPMorgan and Schwab have all, in different ways, said no to agents acting on their customers’ behalf. Their reasons differ — Amazon wants its own agent, the airlines want control of the booking, the banks want control of the money — but the effect is the same: the agent economy has a gate, and the gatekeepers are the companies that already own the customer. OpenAI’s playbook can help a small retailer that blocks agents by accident. It cannot force a platform that blocks agents on purpose to accept them. That means the future of agent shopping will not be decided by which AI company builds the best agent. It will be decided by negotiation, regulation and, eventually, by whether customers demand agent access loudly enough that blocking it becomes more expensive than allowing it.

OpenAI learned the right lesson from its first failure

The contrast between Instant Checkout and ChatGPT Wallet is the most encouraging thing about this launch. The first attempt tried to own the checkout, take a commission and keep the customer inside ChatGPT. The merchants said no, and OpenAI listened. The second attempt leaves the checkout, the customer relationship and the economics with the merchant, and asks only to be the agent that brings the order. That is a much humbler ambition, and a much more realistic one. It also suggests a company that is learning to be a platform rather than a destination: the value is in being the place people start, not the place they finish. If that discipline holds when the volume arrives, and the temptation to take a cut returns, it will be a sign that OpenAI understands the payments business better than most technology companies that have tried to enter it.

The human check is both the safeguard and the bottleneck

Every purchase through ChatGPT Wallet requires the user to review the items, the total and the card, and approve it with a face or fingerprint scan. That is the feature’s strongest protection and its biggest limitation at the same time. It protects against the agent buying the wrong thing, and it protects OpenAI against the liability of an agent spending someone else’s money. But it also means the wallet is not yet what its name promises. A wallet that needs its owner to open it for every single purchase is a remote control, not an agent. The moment OpenAI allows full delegation, with spending limits and pre-approval, the risk changes completely: the agent will be spending money while the user is not looking, and the first time it buys the wrong thing at scale, the argument will not be about technology. It will be about who pays. How OpenAI handles that transition, and how clearly it assigns responsibility when an agent errs, will decide whether the wallet becomes a utility or a cautionary tale.

What to Watch Next

1. How fast the rollout widens. OpenAI has given no date for expanding beyond US Pro subscribers. Watch for the first move to Plus subscribers, to business plans, and to countries outside the United States, where dots themselves are not yet available in the UK or the European Economic Area.
2. Whether Amazon blinks. The largest retailer on the internet currently blocks agent purchases. If Amazon builds an agent checkout of its own, or quietly starts accepting verified agents, the economics of the whole field change overnight.
3. The airlines’ next step. United said it will block agents from booking tickets, and Delta already bans them. Watch whether the block extends to changes, cancellations and refunds, and whether regulators take an interest in who is allowed to buy a ticket.
4. The first full-delegation rules. OpenAI says users will eventually be able to pre-approve purchases within spending limits, categories and time periods. The detail of those safeguards, and who is liable when an agent exceeds them, will be the most important document OpenAI publishes on this product.
5. Fraud and decline rates. OpenAI warns that risk systems may flag legitimate agent purchases. Watch for the first data on how often agent purchases are declined, and whether the fraud firms and payment providers named in the launch manage to recognise them reliably.
6. The banks’ response. American Express is testing in a contained environment, and Chase cards are listed as eligible. Watch for other issuers to set their own agent rules, and for the first dispute in which a customer argues an agent bought something they did not authorise.
7. Whether the networks or the AI companies set the standard. Visa and Mastercard built the rails first, but OpenAI, Google, Meta and Anthropic are building the agents. Watch for a shared standard for agent payments, and for the first sign of a fee, because the company that sets the standard will eventually set the price.

Sources

• OpenAI — Introducing ChatGPT Wallet, 13 November 2026
• Reuters, via WIXX — OpenAI launches ChatGPT Wallet, letting AI agents pay for purchases, 13 November 2026
• Unite.AI — OpenAI Launches ChatGPT Wallet for Agent-Assisted Purchases, 13 November 2026
• The Outpost — OpenAI launches ChatGPT Wallet for agent shopping, 13 November 2026
• PPC Land — OpenAI hands its AI agents a wallet without the real card number in it, 13 November 2026
• Sanbi — Agentic Commerce Q4 2026: Who Can Check Out + Holiday Prep, October 2026