13 min read
OpenAI released ChatGPT Wallet on November 13. It lets the agents running inside ChatGPT pay for things with the user’s credit card, and it was built with American Express, Mastercard and Visa[1]. An agent here means an AI program that takes one goal from a person and works through the steps on its own.
It opened to some ChatGPT Pro subscribers in the US who are 18 or older and hold a US-issued card, and OpenAI said it is starting small to learn from real-world use[2][3]. The agents doing the buying are Dots, the always-on agents OpenAI introduced at its DevDay developer conference in September.
In the same week, United Airlines said it would stop personal AI agents from booking its flights[4], and Charles Schwab and JPMorgan Chase said they are blocking outside agents from logging into customer accounts[5]. Three days separated the launch that hands agents a credit card from the policies that keep agents out.
A payment token stands in for the card number
A user adds a card once. When the agent brings back a product, the person checks the items, the total and the card, then approves with a face scan or a fingerprint. Each approval leaves a record of what the agent was allowed to buy[2][3].
After approval, the card network issues a payment token. A token is a stand-in for the real card that carries its own number, security code and expiry date[1]. Neither the merchant nor the agent sees the actual card details[3]. Tap-to-pay wallets have used the same tokenization approach for years.
Points and cashback accrue as usual, and the right to dispute a defective item with the issuer stays intact, OpenAI said[1][2]. Pricing, delivery, returns and customer service stay with the merchant[3].
OpenAI has already announced a version with fewer approvals. “Over time, we expect users to be able to delegate purchases entirely to their agent by setting clear parameters, such as spending limits, purchase categories, or time periods, and pre-approve those purchases with enhanced safeguards,” the company said[2]. Inside those limits, the agent would buy without a biometric check each time.
Card networks and Walmart allowed agent payments with conditions
OpenAI is not taking a cut of transactions for now, saying it wants to build the relationships with merchants and networks first[1]. Chaya Nayak, who leads monetization and payments at OpenAI, told CNBC: “Our goal here is to really focus on merchants and our network partners and start building out a really great foundation that’s safe and easy to use”[3].
Most of the groundwork predates the launch. Visa and Mastercard have run programs for AI-initiated transactions since 2025, and Google released its Agent Payments Protocol (AP2) the same year[11]. These specify how an agent proves who it is and what the user authorized it to do, and ChatGPT Wallet uses that machinery.
OpenAI published a Commerce Partner Playbook on the same day, a guide for businesses covering how to let ChatGPT agents through a firewall, how to upload product feeds and how to set up a business profile inside ChatGPT[2][3]. It also lined up the payment providers Adyen, Checkout.com and Stripe and the fraud firms Forter, Riskified, Sardine and Signifyd, so that legitimate agent purchases are not declined as fraud[3].
Walmart allows agents to shop through dedicated agent profiles that carry their own spending limits[3]. The agent gets an account of its own instead of borrowing the customer’s.
Airlines and banks blocked agent logins
United Airlines said this week it will restrict personal AI agents from booking and managing reservations on its digital platforms. The airline said that “sharing account credentials with personal, third-party, AI assistants can expose sensitive information, and actions taken by these tools may result in unintended purchases, account changes, or use of MileagePlus miles”[4]. It also cited federal aviation rules. A human has to confirm at check-in that their bags contain no prohibited items, and an agent cannot make that attestation[4].
The restriction reaches past United’s own app and website. “As a result, United is taking steps to restrict access, directly within our digital platforms and indirectly via travel agencies and aggregators, to prohibit use by unauthorized, third-party personal AI agents or assistants. This work includes updates to our booking, ticketing and distribution policies,” the airline said[4].
Banks and brokerages started blocking agents in October. Charles Schwab chief executive Rick Wurster told advisors at the firm’s conference in Boston on October 28: “The first thing we are doing is we are keeping rogue agents from being able to log in on the client’s behalf, we block that activity using technology.” He added: “When you look on a weekly basis, it’s 30, 40, 50 times that we might be seeing a particular AI agent come to our sites and try to log in, and we are blocking it”[5]. Schwab is building its own assistant, Charley, and controlled connections instead[5].
JPMorgan Chase said on its October 13 earnings call that it will not let outside agents operate inside customer accounts, and a spokeswoman confirmed to CNBC that the bank currently blocks external AI agents acting for retail customers. Citigroup and Bank of America have imposed similar limits, the Financial Times reported[5]. Amazon blocked Meta’s shopping agent Muse from buying on its site[2][3].
Agent identification separates what gets through from what gets blocked

The two sets of announcements appear to cover the same ground, yet the blocked behavior and the permitted behavior do not overlap.
Banks and airlines are blocking agents that borrow a human account. The user hands over a username and password, and the AI logs in as that person and clicks through the site the way a human would. The bank has no way to tell whether the session belongs to the customer or to a program, and the password now sits inside an outside service. Wurster put it this way: “The good news that we’ve found actually is it doesn’t seem that a lot of clients have a lot of interest in giving up their credentials and their financial experience to these agents”[5].
ChatGPT Wallet works the other way. The agent identifies itself and comes in through a channel the card network created for it, and what the user authorized is recorded at each approval[2][3]. Walmart’s agent profiles separate the traffic the same way[3].
Virtual cards fall somewhere between the two. Agent shopping tools often generate a single-use virtual card to pay with, and JPMorgan classifies those as commercial cards, which is causing some agent purchases to be declined[6]. That is also why ChatGPT Wallet does not accept virtual cards at launch.
So the rule taking shape decides which agents get in, not whether agents get in at all. An agent that declares itself goes through, and an agent hiding behind a customer’s login does not.
Visa measured shopping agents and found a 57.2% pass rate
Visa published the first results of its Agentic Commerce Benchmark on November 5. Agentic commerce means a purchase where an AI agent, not a person, picks the product and completes the transaction. The test put 22 models from 12 labs through 15 shopping tasks[7]. A lab in this context is a company that trains large AI models of its own, such as OpenAI or Anthropic.

The overall pass rate was 57.2%. The best model passed 82.2% of its runs and the worst 28.9%[7]. The tasks involved buying a TV, a laptop or a personal care appliance, with the customer withholding some of their requirements. The agent had to draw those out in conversation, and if nothing on offer fit, it had to decide not to buy[7].
In 54% of the failed runs, the agent bought something anyway. It either picked a product that did not meet the customer’s needs or moved to checkout before it had gathered all of them[7]. When the same scenario was run three times, all three runs passed only 34.2% of the time. Models often failed a scenario they had already passed[7].
The five models that uncovered the most hidden requirements passed 70.2% of their runs, against 44.4% for the bottom five. Cost tracked performance too, though less sharply: 67.1% for the five priciest models per run and 50.2% for the five cheapest[7]. This first round covered requirement discovery and the buy-or-decline decision. Actually submitting payment and completing checkout is not in scope yet[7].
Rubail Birwadker, Visa’s global head of growth, said of the results: “AI agents have the potential to power the next era of commerce, but trust is key to making it work.” He told Fortune: “We’re not there yet from a trust or technology perspective”[7]. In Visa’s Trust Index published in September, 72% of US consumers had used an AI assistant, while 23% said they trusted AI to make payments for them[7].
Juniper Research forecasts that agentic commerce will grow from roughly 300 million users in 2026 to 1.3 billion by 2031, with spending through agents reaching $1.5 trillion by 2030. The biggest obstacle named in the same report is consumer trust[8].
Outside the US, card issuers are still agreeing on the format
ChatGPT Wallet does not work with cards issued outside the US, but similar pilots started earlier elsewhere. In South Korea, Shinhan Card said on March 30 that it had completed the country’s first live agent-paid transaction with Mastercard. An agent found and booked a ride through a global mobility service and the payment went through with a single user approval. The two companies designed the login checks, the permission handling, the payment flow and the merchant hookup together[9].
KB Kookmin Card is working with a commerce AI vendor to test what an agent needs to clear a merchant site, a payment module and a card network in one transaction, including identity verification and transaction checks. Shinhan Card and SK Telecom have begun drafting a standard specification for agent payments[10].
Korean checkout still has steps a human has to tap through. Paying inside an app usually means entering a PIN set for that wallet, and opening an account or raising a limit often means confirming your identity through your mobile carrier. Whether an agent may press those on a user’s behalf has no published regulator or issuer standard that I could find. What is public so far stops at the card companies agreeing on a format[10].
What to actually do
- Do not hand a username and password to an AI tool. That is precisely the pattern banks and airlines are blocking, and United said it revised its booking, ticketing and distribution policies to restrict it[4]. Credentials stored in an outside service widen the blast radius of any breach to the whole account.
- Keep per-purchase approval on when an agent pays. Pre-approved limits and categories are something OpenAI has only announced as a next step[2].
- Check what the agent bought against what you asked for. In Visa’s test, agents pushed through to a purchase in 54% of failed runs[7]. Checking after the charge clears means paying return shipping.
- Remember that returns and refunds run through the merchant. With ChatGPT Wallet, pricing, delivery, returns and support stay with the business, and card dispute rights are unchanged[1][3].
- Upgrading a plan outside the US will not turn this on. The card itself has to be US-issued[2].
One more thought
Building the payment method and accepting the payment are jobs for different companies. OpenAI and three card networks can agree on a token format, and the transaction still fails if airlines and banks decline it. Both decisions landed in the same week.
Visa put its name on ChatGPT Wallet in the same month it published its own measurement showing a 57.2% pass rate for shopping agents[7]. The company that helped build the plumbing is also the one saying the agents are not ready.
The consumer research points the same way. 72% of US consumers have used an AI assistant and 23% would let one pay[7]. What closes a 49-point gap like that is probably not another feature. It is an agreed answer to who pays when the agent gets it wrong, which is close to the reason the banks give for blocking.
Sources
- OpenAI unveils ChatGPT Wallet to enable purchases through AI agents (Anadolu Agency)
- OpenAI rolls out ChatGPT Wallet to let Dots AI agents make purchases (Android Authority)
- OpenAI’s ChatGPT Wallet lets its AI agents shop without seeing your card (The Next Web)
- United Airlines blocks AI agents from booking flights, citing cybersecurity risks (TechCrunch)
- Schwab CEO says firm is blocking rogue AI agents from client accounts (InvestmentNews)
- Visa tested 22 AI models as shoppers, and they failed 42.8% of the time (The Next Web)
Notes
- Anadolu Agency (published 2026-11-13, checked 2026-11-14). The November 13 announcement, joint development with American Express, Mastercard and Visa, one-time card entry with a security check per transaction, agents having no access to real card details while networks issue transaction-specific tokens carrying a limited-use card number, security code and expiry date, rewards and protections preserved, no initial cut of transactions, and the history of Instant Checkout from September 2025 OpenAI unveils ChatGPT Wallet to enable purchases through AI agents ↩
- Android Authority (published 2026-11-13, checked 2026-11-14). Eligibility (US-based ChatGPT Pro subscribers with a US-issued Visa, Mastercard or American Express), the purchase flow (review items, total and card, then approve with Face ID or a fingerprint), OpenAI’s quoted statement on delegating purchases with spending limits, purchase categories or time periods, the Commerce Partner Playbook, Amazon blocking Meta’s Muse as reported by GeekWire, and United’s policy OpenAI rolls out ChatGPT Wallet to let Dots AI agents make purchases ↩
- The Next Web (published 2026-11-13, checked 2026-11-14). The 18-and-over condition, “starting small to learn from real-world use”, merchants and agents not seeing real card details, pricing, delivery, returns and support staying with the business, Chaya Nayak’s quote via CNBC, American Express’s Luke Gebb, the fraud firms Forter, Riskified, Sardine and Signifyd and payment providers Adyen, Checkout.com and Stripe, Walmart’s agent profiles with spending limits, and Schwab and JPMorgan blocking money movement OpenAI’s ChatGPT Wallet lets its AI agents shop without seeing your card ↩
- TechCrunch (published 2026-11-12, checked 2026-11-14). United Airlines’ statement in full on credential sharing and on restricting access directly and via travel agencies and aggregators, plus the federal rule requiring a human to attest to bag contents at check-in. The Next Web dated the announcement to Wednesday, November 11; TechCrunch published on November 12 United Airlines blocks AI agents from booking flights, citing cybersecurity risks ↩
- InvestmentNews (published 2026-10-28, checked 2026-11-14). Rick Wurster’s keynote quotes at Schwab IMPACT 2026 on blocking rogue agent logins and the 30-to-50-times-a-week figure, his October 15 earnings call line on not allowing agents to move money, Schwab’s Charley assistant, Jamie Dimon’s October 13 earnings call position and the JPMorgan spokeswoman’s confirmation via CNBC, similar limits at Citigroup and Bank of America per the Financial Times, and Robinhood’s opposite choice Schwab CEO says firm is blocking rogue AI agents from client accounts ↩
- Crypto Briefing (published 2026-11-12, checked 2026-11-14). JPMorgan classifying the single-use virtual cards agents often use as commercial cards, causing some purchases to be declined, with context from Dimon’s October 13 earnings call JPMorgan blocks some AI agent purchases on Chase cards as merchants raise concerns ↩
- The Next Web (published 2026-11-05, checked 2026-11-14) reporting the first round of Visa’s Agentic Commerce Benchmark. 22 models from 12 labs, 15 shopping tasks, 57.2% overall pass rate (42.8% failure), best 82.2% and worst 28.9%, 34.2% passing all three attempts, purchases completed in 54% of failed runs, 70.2% for the top five requirement-discovery models against 44.4% for the bottom five, 67.1% for the five most expensive models against 50.2% for the five cheapest, and checkout excluded from this round. Rubail Birwadker’s quotes and Visa’s September Trust Index figures of 72% assistant usage and 23% trust in AI payments Visa tested 22 AI models as shoppers, and they failed 42.8% of the time ↩
- FStech (published 2026-06-29, checked 2026-11-14) reporting Juniper Research. Agentic commerce growing from about 300 million users in 2026 to 1.3 billion by 2031, spending through agents reaching $1.5 trillion by 2030, traditional checkouts persisting alongside, and consumer trust as the biggest barrier Agentic commerce to grow 350% by 2031, finds Juniper Research ↩
- Asia Economy (published 2026-03-30, checked 2026-11-14). Shinhan Card completing the first live AI agent payment among Korean card issuers with Mastercard, integrated with a global mobility service, completed with a single user approval, with authentication, permission handling, payment process design and merchant integration designed jointly Shinhan Card completes Korea’s first live ‘AI agent pay’ transaction ↩
- Newsway (published 2026-10-03, checked 2026-11-14). KB Kookmin Card’s agent payment pilot with SY Polaris covering the path from merchant site through payment module to card network, including identity verification and transaction checks, and Shinhan Card’s work with SK Telecom on a standard specification for agent payments AI shopping era nears as card industry pushes into next-generation payments ↩
- Cointelegraph (published 2026-11-13, checked 2026-11-14). Google introducing its Agent Payments Protocol in 2025 and Visa and Mastercard running initiatives for AI-initiated transactions, plus the biometric approval and tokenization structure, the US Pro rollout, and the plan for preset spending limits OpenAI launches AI payments in ChatGPT with Visa, Mastercard and Amex ↩