OpenAI Hands Its AI Agents a Wallet as Google Takes the Opposite Bet on Your Money
On Friday, 13 November 2026, the two companies that dominate the consumer artificial intelligence market made the same announcement on the same day — and gave opposite answers. OpenAI launched ChatGPT Wallet, a feature that lets its AI agents pay for things on a user’s behalf. Google connected its Gemini assistant to Google Wallet, letting it read a user’s spending history and offer financial advice. One handed its AI the power to spend money. The other handed its AI the power to look at money, and nothing more.
The timing was not a coincidence of the calendar. Both launches land in the middle of a race to turn chatbots into agents — software that does not merely answer questions but carries out tasks in the real world. Shopping is the most obvious task of all, and the most sensitive: it involves real money, real merchants and real liability. How the two companies have chosen to handle that sensitivity now defines two competing visions of what an AI assistant is for, and the contest between them will shape how hundreds of millions of people buy things online in the years ahead.
OpenAI’s move is the bolder of the two. According to the company’s announcement, ChatGPT Wallet allows its always-on agents, known as dots, to complete purchases for users, from finding a product to paying for it. The feature was built with American Express, Mastercard and Visa, and it is rolling out first to ChatGPT Pro subscribers in the United States who are aged 18 or older and hold eligible US-issued cards, according to CNBC. Every purchase still requires the user’s approval, confirmed with a security check such as Face ID or a fingerprint scan. What changes is who does the work: the agent browses, chooses and checks out, and the human signs off.
Google’s move, announced the same day, is deliberately narrower. Gemini can now connect to Google Wallet through a feature called Personal Intelligence, which links information from everyday apps such as Gmail, Photos, YouTube and Search. Users type “@Google Wallet” in the chat to ask questions about their spending. According to Google’s announcement, Gemini can only read account information. It cannot move funds, manage money or invest. The connection is off by default, and users choose to switch it on in the Personal Intelligence settings.
The contrast is the story. One company believes the future of AI commerce is delegation: give the agent the job and let it finish it. The other believes the future is advice: give the agent the information and let the human decide. Both launched on the same Friday, both aimed at the same wallet in the same pocket, and both cannot be right about what users will want.
How ChatGPT Wallet works
ChatGPT Wallet lives inside dots, the always-on agents OpenAI introduced at its developer conference on 29 September 2026. Dots run on their own cloud computer with their own browser, and they can work through tasks — calendars, travel booking, everyday errands — while the user does something else. Until now, a dot could find something to buy but could not pay for it. Wallet closes that gap.
The mechanism is built around a simple idea: the agent should never see the card. According to OpenAI, a user enters their card details once. When a dot finds something to buy, the user reviews the item, the total amount and the card to be used, and then approves the purchase with a security check, such as Face ID or a fingerprint. Only then does the payment network — American Express, Mastercard or Visa — issue a payment token: a substitute card number with its own security code and expiry date, valid for that single transaction. The merchant and the agent receive the token, never the real card details. The charge is billed to the selected card exactly as an ordinary purchase would be.
OpenAI describes Wallet as the first wallet that lets agents pay without access to a user’s actual card details while requiring a security check for every purchase. The company says cardholders keep their existing rewards and protections, and that requests for refunds or disputes over an order go to the merchant. OpenAI is neither the bank nor the card issuer, the payment network, the lender nor the merchant for a wallet purchase, according to terms the company published on 12 November, as reported by Quartz. OpenAI does not hold the purchase funds.
The launch is deliberately small. It is available only to ChatGPT Pro subscribers in the United States who are 18 or older, with an eligible US-issued card. According to OpenAI’s help centre, as reported by several outlets, eligible cards are American Express, Chase, Mastercard and Visa credit cards with a US billing address; debit cards and virtual cards are not supported at launch. The wallet works only in dots conversations, not in ordinary ChatGPT chats, and OpenAI has given no date for a wider rollout or for other countries.
OpenAI is also taking no share of transactions at launch. Chaya Nayak, who leads product for monetisation and payments at the company, told CNBC that the work was about building relationships with merchants and the payment networks. Luke Gebb, American Express’s executive vice-president and global head of innovation, said the restricted launch was a way to check that the feature worked before a broader rollout. Sherri Haymond, Mastercard’s global head of digital commercialisation, was also quoted in coverage of the launch. The message from all three companies is the same: this is a pilot, and the scale will follow the evidence.
What comes next is already sketched out. OpenAI says it expects users eventually to be able to give agents greater authority over purchases by setting limits on spending, categories or time periods, with enhanced safeguards that build on biometric checks and authorisation records. At launch, though, no purchase can complete without the user’s review and security check. The delegation is real, but it is not yet autonomous.
Alongside the wallet, OpenAI published a Commerce Partner Playbook, guidance telling businesses how to prepare for agent shoppers. It covers claiming and verifying a business in ChatGPT, supplying product information and prices, making sure a website does not inadvertently block an agent’s browser, and reducing unnecessary payment rejections by coordinating with payment providers. The company names Forter, Riskified, Sardine and Signifyd as fraud partners, and Adyen, Checkout.com and Stripe as payments partners. The playbook’s existence is itself a signal: OpenAI knows the wallet only works if the rest of the internet lets its agents in.
A second attempt at the checkout
ChatGPT Wallet is OpenAI’s second attempt at putting a payment step inside ChatGPT. The first, Instant Checkout, launched in September 2025 with Stripe and sellers on Etsy in the United States, letting people buy from participating merchants without leaving the chat. By March 2026, OpenAI had moved away from it, replacing it with a revised shopping experience built around finding and comparing products. Nayak told CNBC that Wallet builds on that product-discovery work: once someone has found something to buy, the agent can now help carry the purchase through to payment.
The difference between the two attempts matters. Instant Checkout was a checkout button inside a chat. Wallet is a payment credential for an agent that acts on its own. The first asked the user to buy something they had found. The second asks the user to let software buy something on their behalf. That is a larger claim about trust, and OpenAI has spent the year building the infrastructure for it: dots in September, verified business profiles in ChatGPT in beta on 10 November, and now the wallet itself.
The scale of the audience makes the stakes plain. According to CNBC, ChatGPT has more than 1.2 billion weekly active users and nearly 100 million paying users as of the end of the third quarter of 2026. Even a feature limited to Pro subscribers in one country reaches millions of people. If agents become a normal way to shop, the volume of transactions they generate could become one of the largest new payment flows in the world — and the companies that control the rails will want a say in how it works.
Google’s answer: look, don’t touch
Google’s launch on the same day could hardly be more different in design. The Gemini app in the United States can now connect to Google Wallet, rolling out sequentially across the Android, iOS and web versions. The integration sits inside Personal Intelligence, the feature that links information from a user’s everyday apps to personalise responses. Google Wallet has now been added as a connection point alongside Gmail, Photos, YouTube and Search.
What Gemini does with the connection is analysis, not action. In response to a request such as reviewing last quarter’s spending and setting a savings goal for the next three months, it generates charts within the chat showing cash flow, asset allocation and budget trends. It can identify recurring charges from linked accounts and notify the user if a price has gone up. For holiday-season budgeting, it makes proposals based on the previous year’s gift and travel spending. If users optionally link credit cards or bank accounts through the financial technology company Plaid, more in-depth analysis becomes possible.
The guardrails are the point. Google explains that users themselves choose to connect through “Connected Apps” in the Personal Intelligence settings, and the connection can be toggled on and off at any time. Gemini can only read account information and cannot engage in any way with moving or managing funds or investing. Google has also added a “Memory Summary” that lets users check and edit what the AI remembers, displaying saved preferences, interests and ongoing goals in one place.
The background to Google’s caution is, by most accounts, the sensitivity of the data itself. Transaction history reveals asset status and lifestyle habits, and once an AI gains write access to money, damage from an erroneous operation or misuse would be difficult to reverse. By keeping connections off by default and limiting the assistant to reading, Google has made clear that it is prioritising safety over convenience — at least for now.
Google’s timing also fits a wider push. The company introduced Gemini 4 Hydrogen, its fast execution model, on 12 November, the day before the Wallet integration, and on Monday 16 November it expanded its scientist access programme to include Gemini 4 Carbon, giving 10,000 scientists six months of free access to research tools. The Wallet connection is one piece of a broader effort to make Gemini useful in everyday life, not merely impressive in a demonstration.
The shops are not all saying yes
For all the careful design, ChatGPT Wallet faces a problem that no amount of tokenisation can solve: many of the internet’s biggest shops do not want AI agents buying things on their platforms at all. United Airlines said on 11 November that it would block agents from accessing its systems to book tickets, according to CNBC. Amazon prevents agents from buying products on its platform, CNBC also reported. Two of the largest online sellers in the world have, in effect, closed the door to the agent before it has even arrived.
They are not alone. Blend, a Shopify conversion agency, sent an AI shopping agent to 200 retailers in the United States and the United Kingdom in September 2026. Twenty-nine of them refused to let the agent in at all, even on a second try, including Home Depot, Best Buy and Macy’s. The audit gave the 200 stores an average score of 65 out of 100 for how well an agent could load the site, find products and reach checkout. The experiment measured a real obstacle: many sites are built to keep automated traffic out, and an agent’s cloud browser looks, to a security system, very much like a bot.
This is why OpenAI published the Commerce Partner Playbook, and why it talks about Web Bot Auth — a way for verified agent traffic to identify itself — and about a Forwarded header that relays user signals to a merchant’s systems. The company is asking retailers to distinguish between a legitimate agent acting for a real customer and a scraper or a fraudster. That distinction is technically possible but commercially fraught. A retailer that lets agents in gives up control over how its products are presented, compared and discounted. A retailer that keeps them out risks losing the customer whose agent simply goes elsewhere.
Visa has been measuring the same problem from the other side. The card network’s agentic commerce benchmark put 22 AI models from 12 laboratories through 15 shopping tasks, according to a report published on 5 November. The exercise was designed to find out how well agents actually shop — whether they can complete a purchase accurately, cheaply and without error. The results, and the fact that Visa felt the need to run the test at all, underline how new this territory is. Nobody yet knows how reliable an agent shopper is, because nobody has measured one at scale.
The banks are split, too
The financial institutions whose cards the agents would spend are themselves divided. According to market reports cited by BigGo Finance, Citigroup, Bank of America and Capital One permit authorisation by users for AI payments, while JPMorgan Chase explicitly prohibits AI access to credit cards. The split matters because a wallet that works with some banks’ cards and not others is a wallet whose usefulness depends on which card happens to be in the user’s pocket — and that, in turn, could become a factor in which card a customer chooses.
The card networks, by contrast, are competing to be the infrastructure of agent commerce. American Express, Mastercard and Visa all signed on to ChatGPT Wallet at launch, and all three have been building their own agent payment products. For them, an AI agent is a new kind of user, and every transaction it completes is volume on their rails. The commercial logic is straightforward. The harder question — who bears the loss when an agent buys the wrong thing, or a fraudster persuades an agent to buy the wrong thing — has no settled answer yet, and the networks, the banks, the merchants and OpenAI will all have an interest in how it is settled.
There is also a quieter tension in the background. OpenAI says it takes no cut of wallet transactions at launch. That will not last forever, and everyone involved knows it. The company has nearly 100 million paying users, according to CNBC, and a payment flow running through its agents would eventually be worth a great deal — whether through a fee, through the data, or through the leverage it gives OpenAI over merchants who want access to its users. The zero fee at launch is a strategy, not a principle: it buys adoption now and leaves the business model for later.
Analysis: Why It Matters
The checkout is the new front line of the AI race
For three years, the contest between the AI laboratories has been fought on capabilities: which model is smarter, faster, cheaper. The wallet launches move the contest somewhere else entirely — to the moment money changes hands. That moment is where an assistant stops being a clever tool and becomes an economic actor, and it is where the value in the whole enterprise may eventually concentrate. Whoever controls the checkout controls the relationship with the customer, the data about what was bought, and the fee for making it happen. OpenAI and Google both understand this, which is why both moved on the same day. The race is no longer only about intelligence. It is about permission: who is allowed to act, and on whose behalf.
Tokenisation is the trust mechanism — and its limit
The cleverest part of ChatGPT Wallet is not that an agent can pay, but that it can pay without ever holding the card. Tokenisation — the substitute number, security code and expiry date issued for a single transaction — is a genuine piece of engineering, and it answers the most obvious objection to handing a wallet to software: that the software might leak, lose or misuse the credentials. A token that works once and then dies is a credential that cannot be stolen and reused. But tokenisation solves only the problem of the card. It does not solve the problem of the choice. An agent that buys the wrong size, the wrong model, the wrong quantity, or the right thing at the wrong price has not leaked anything — it has simply made a mistake with real money. The biometric approval is meant to catch that, but it catches it only if the user actually reads what they are approving. The history of cookie banners, software updates and terms-of-service checkboxes suggests that many will not. The safety of the system will depend less on cryptography than on human attention, and human attention is the scarcest resource in the design.
Google’s caution is a product decision, not a technical one
It would be a mistake to read Google’s read-only approach as a lack of ambition. Google could build an agent that pays — it has the payment network, the merchant relationships and the assistant. It has chosen not to, and the choice reveals a different theory of the user. Google’s theory is that people are wary of handing financial data to an AI, and that the fastest way to earn trust is to prove the assistant can be useful without being dangerous. A Gemini that spots a subscription price rise or builds a holiday budget is helpful in a way that is easy to verify and hard to regret. An OpenAI dot that buys something is helpful in a way that is easy to regret and hard to undo. Google is betting that trust accumulates slowly and that the company which earns it first will win the right to do more later. OpenAI is betting that convenience wins first and trust follows. Both theories are plausible. Only one will be proved right, and the evidence will arrive in the form of ordinary people’s behaviour over the next year.
The merchant veto could decide the winner
The most important fact in this story may be the least glamorous: 29 out of 200 retailers would not let an AI shopping agent through the door, and two of the biggest — Amazon and United Airlines — have said no outright. An agent that cannot buy from Amazon is an agent that cannot buy most things, and an agent that cannot book a flight is an agent that cannot plan a trip. OpenAI’s playbook is an admission that the wallet’s success depends on persuading thousands of independent businesses to change how their websites treat automated visitors. That is a negotiation, not a launch, and it will be fought shop by shop. Google’s approach sidesteps the fight entirely: reading a bank statement requires no merchant’s permission. If the merchant resistance holds, the read-only assistant may turn out to be the more useful product — not because it does more, but because it is allowed to do what it does everywhere.
Liability is the question nobody has answered yet
Sooner or later, an agent will buy something its user did not want, or be tricked into buying something nobody wanted, and the argument about who pays will begin. OpenAI’s terms say the merchant remains the seller and that refunds and disputes go to the merchant. The card’s protections remain available for eligible disputes. But those are the rules for human shoppers. An agent sits between the human and the merchant in a way no previous intermediary has: it interprets the instruction, chooses the product and completes the payment. When it gets something wrong, was the instruction unclear, the agent careless, or the merchant misleading? The Federal Reserve’s latest Financial Stability Report found that 87 per cent of the 23 market contacts it surveyed named artificial intelligence as a risk, up from 50 per cent in the spring, according to a roundup of the report — a sign that the financial system is already watching this question. The law has not caught up with agentic commerce, and the first expensive mistake will do more to define it than any launch announcement. Until liability is settled, every wallet launch is also a legal experiment, and the users are the test subjects.
What to Watch Next
- How many Pro subscribers actually use Wallet in its first month, and whether OpenAI publishes any usage figures — the gap between a launch and a habit is where most products die.
- Whether Amazon and United Airlines soften their blocks, or whether other large retailers follow them — the merchant map will decide how useful an agent shopper can be.
- Whether JPMorgan Chase maintains its prohibition on AI access to credit cards, and whether other banks join it or break ranks — a bank split could fragment the wallet by card.
- When OpenAI introduces the promised spending limits, purchase categories and time periods for pre-approved purchases — the moment delegation becomes genuinely autonomous.
- Whether Google extends Gemini beyond read-only access, and how quickly — any move by Google to let Gemini pay would signal that the cautious bet has changed.
- What fee, if any, OpenAI begins to charge on wallet transactions, and how merchants and networks react — the zero per cent launch rate is a strategy with an expiry date.
- The first public dispute over an agent purchase that went wrong, and who absorbs the loss — the merchant, the network, the bank or the user — because that case will set the precedent for the whole market.
Sources
- OpenAI, “Introducing ChatGPT Wallet”, 13 November 2026
- BigGo Finance, “AI Makes Full-Scale Push into Household Finances and Payments”, 16 November 2026
- AI Affairs, “OpenAI launches ChatGPT Wallet for purchases made by its dots agents”, 13 November 2026 (reporting CNBC)
- Decrypt, “OpenAI Launches ChatGPT Wallet to Let AI Agents Shop — Without Crypto”, 14 November 2026
- Unite.AI, “OpenAI Launches ChatGPT Wallet for Agent-Assisted Purchases”, 13 November 2026
- Silicon Report, “OpenAI announces ChatGPT Wallet to let AI agents make purchases”, 13 November 2026