The Wallet Is Live. The Chargeback Rules Are Not.

OpenAI’s ChatGPT Wallet shipped on 13 November 2026 with Visa, Mastercard and Amex tokens; the dispute evidence merchants need for agent orders does not exist yet

On Friday 13 November, OpenAI turned on ChatGPT Wallet. A US subscriber on the Pro plan, 18 or older, can now add a card issued by Visa, Mastercard or American Express, ask one of OpenAI’s shopping agents to find something, and approve the purchase with Face ID or a fingerprint. The card number never reaches the agent. The network issues a token scoped to that transaction, and the merchant receives a shipping address and a confirmation.

The launch is small on purpose. It covers US Pro subscribers only, on US-issued cards, and OpenAI says it is not taking a cut of transactions to start. Chaya Nayak, who leads product for 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.”

Everything in that paragraph works. The part that does not work yet is the argument that happens three weeks later, when a customer looks at a statement, does not remember ordering a $340 blender, and calls the issuer.

A token proves who paid, not what was agreed

Payment tokens are a well understood piece of plumbing. Visa and Mastercard have issued them for years to keep card numbers out of merchant databases. A token scoped to one merchant, one amount and one agent is a modest extension of the same idea, and the networks have built exactly that: Mastercard’s agentic tokens, Stripe’s shared payment tokens, Amex’s agent programme.

The token answers a narrow question. It proves that a card was presented by an agent that the issuer had registered, acting for a cardholder who authenticated with a passkey or a biometric. That is a real improvement over what agent commerce looked like a year ago, when assistants were simply given a password and told to log in.

It does not answer the question a dispute actually turns on. Did the agent buy what the cardholder asked for? A human checkout leaves a trail: a browser session, a device fingerprint, a cart the person filled, a delivery address the person typed. An agent checkout leaves a token, a mandate and a model. The signals that a merchant normally uses to defend a chargeback are the exact signals that the agent exists to skip.

That is not a technology gap that another release closes. It is a gap in the record of intent.

No network has a code for this

Ask the card networks what happens when an agent order is disputed, and the honest answer as of November 2026 is that nothing specific happens. There is no dispute reason code for an agent-initiated purchase. Worldpay’s July 2026 analysis found no formal liability shift for agent purchases, which means the transaction follows ordinary card-not-present rules. Visa and Mastercard have both published trust frameworks. Neither has published a rule that moves the loss.

What they have published is verification machinery:

  • Visa Intelligent Commerce, live since July 2026 with more than 30 European issuers and merchants including lastminute.com, Frasers and Cleverbridge, built around the Trusted Agent Protocol that lets a merchant recognise a registered agent before it buys.
  • Mastercard Agent Pay, which issues agentic tokens to registered agents and, with Google, announced Verifiable Intent in March 2026 to create a cryptographic record linking the cardholder, the instructions and the transaction.
  • EMVCo’s draft framework for card-based agent payments, published on 1 September 2026, built around what it calls Intent Services: a shared record of what the consumer authorised that persists before, during and after a transaction.

Read that list carefully. Every item is about identity and authorisation. None of them is a chargeback rule, and none of them creates a new place to file the loss.

The practical consequence is that agent orders land in the same bucket as any card-not-present sale, with less evidence than usual. A merchant’s dispute ratio is calculated across all sales, so a new stream of poorly evidenced transactions shows up in a number that acquirers watch. The volume is small today. The direction is not reassuring. Nobody has published what share of agent orders end in a dispute, which is why every number in this market is a projection built on a sample nobody outside the networks can see.

What a merchant actually holds after the sale

The evidence playbook for agent orders is being written by the people who process the disputes, and it is specific. A merchant defending an agent chargeback needs the agent platform’s identity for that session, the mandate showing what the cardholder authorised, a comparison of the order against the customer’s stated limits, a timestamp showing the customer was notified and could cancel, and proof of delivery to the customer’s address.

Most merchants have none of those fields in their order record. They have a token reference and a shipping address.

The workaround that is circulating among payment teams is to treat an agent order as a delayed-capture transaction. Authorise the card, hold the capture for 15 to 60 minutes, and push a notification to the human: your agent just ordered this, tap here to void it. A void on an unsettled authorisation costs nothing. A refund after capture costs a fee, and a chargeback costs the fee plus the goods in many cases.

That design has a cost of its own. Holding capture means holding inventory, and it slows down the one thing agents are supposed to speed up. It also assumes the cardholder reads the notification, which is the same assumption that failed for every fraud alert of the last decade.

None of this makes agent commerce unworkable. It makes it a merchant operations project before it is a consumer habit.

The 1.5% that everybody quotes

The forecasts for this market disagree by a factor of fifty, and the disagreement is almost entirely about definitions. That is worth knowing before anyone plans against a number.

EMARKETER puts US retail ecommerce running through AI platforms at about $20.57 billion in 2026, roughly 1.5% of the total. Morgan Stanley’s base case has agentic shoppers at about 1% of US ecommerce spending in 2026, rising to 10% by 2030, which it values at $190 billion to $385 billion. Bain counts purchases that agents initiate, influence or complete, and lands at $300 billion to $500 billion. McKinsey counts revenue that agents orchestrate, including discovery, and gets $900 billion to $1 trillion in the US and $3 trillion to $5 trillion globally.

Table 1: 2030 forecasts for agentic commerce and what each firm counts

Firm 2030 US estimate Share of US ecommerce What the number counts
Morgan Stanley $190bn to $385bn 10% to 20% Spending by shoppers who use agents
Bain $300bn to $500bn 15% to 25% Purchases agents start, sway or finish
McKinsey $900bn to $1tn not stated Revenue agents help orchestrate, goods only
EMARKETER $20.57bn in 2026 1.5% in 2026 Sales completed through AI platforms

Source: Morgan Stanley, Bain and McKinsey forecasts as compiled by Retention Edge, with EMARKETER’s 2026 estimate from the same compilation

The spread is not a failure of research. It is what happens when four firms count four different things, and the widest definition wins the headline. A merchant reading “one trillion dollars by 2030” and a merchant reading “1.5% of ecommerce today” are looking at the same market through different lenses.

The measured behaviour sits closer to the small number. A Visa survey conducted by Harris Poll in September 2026 found that 72% of US consumers had used an AI assistant, but only 23% trusted generative AI to handle a payment on their behalf. Accenture’s January 2026 survey of 25,590 consumers across 16 countries found 74% would let an agent handle routine tasks and 32% would let it choose what to buy, but only 9% were open to a fully autonomous purchase.

People will let an agent do the research. They are not yet ready to let it spend.

Where the agent checkout already works

The counter-example sits in China, and it is worth studying because it removes the exact variable that is blocking the US market. Alibaba’s Qwen assistant handled more than 120 million orders during six days of the 2026 Lunar New Year shopping period. That volume did not come from a new payment rail. It came from an assistant embedded in a platform that already held the customer’s payment credential, the merchant’s catalogue and the dispute process.

When one company owns discovery, checkout and the refund, an agent order has no liability seam to argue about. The platform can reverse a purchase in the app, and the customer never calls an issuer. The cost of that convenience is concentration, which is a different problem from the one OpenAI is solving.

Western markets have split the roles. The assistant is OpenAI’s, the credential is a bank’s, the merchant is someone else’s, and the dispute is a network’s. Each party can verify its own step. None of them owns the whole transaction.

The conversion data hints at what is being given up. Adobe Analytics found that AI-referred traffic to US retail sites converted 42% better than other channels in March 2026, after converting 38% worse a year earlier. Traffic that arrives with a stated intent is worth more than traffic that arrives from a banner. That is the prize, and it explains why merchants keep testing agents even while their fraud teams object.

The adoption numbers inside the assistants are just as lopsided. Meta’s Muse passed 10 million users, with AI-assisted shopping among its most used features, while DoorDash and Uber have reported low numbers of orders placed by agents. People use assistants to decide. They still click the buy button themselves.

Two blocks and one withdrawal

The resistance is not theoretical. It arrived in the same week as the wallet.

On 12 November, United Airlines said it would restrict AI agents from booking flights and managing reservations on a customer’s behalf, citing cybersecurity risk and the federal rule that requires a human to complete check-in and confirm that bags contain no banned substances. The airline said it would block unauthorised third-party agents directly on its platforms and indirectly through travel agencies and aggregators.

The week before, Bank of America told customers not to share their passwords with AI assistants. Both moves point at the same hole: an agent that logs in as the customer is indistinguishable from the customer, until something goes wrong.

Then there is OpenAI’s own history. The company launched Instant Checkout in the autumn of 2025 and charged merchants 4% of the transaction. By March 2026 it had stepped back from in-chat checkout and moved purchases to the merchant’s own site. Walmart’s Daniel Danker told Wired that conversion was three times lower for items bought inside ChatGPT’s checkout than for items that sent shoppers to Walmart’s own site.

The wallet’s zero take rate is the response to that experiment. OpenAI is not monetising the transaction. It is buying the right to be the place where the transaction starts, and paying for it with margin it could have taken. Nayak’s framing, that the focus is merchants and network partners, is the language of a company rebuilding trust it spent once already.

Zero take rate is a price, not a gift

Work out what the zero costs. If every dollar of the $20.57 billion that EMARKETER expects to flow through AI platforms in 2026 carried a 4% platform fee, the annual revenue at stake would be about $823 million. OpenAI has chosen to forgo that for now, in exchange for merchant relationships and a position in the checkout flow.

That is a rational trade for a company that already lost a year to a checkout product merchants did not want. It is also a warning for anyone building on top of the wallet. A platform that is not charging for payments will charge somewhere else, and the likely places are advertising, product placement and the data that comes back from the order.

The fee stack explains why merchants are cautious rather than enthusiastic. A 4% platform fee does not replace the card fees a merchant already pays. It sits on top of them. For a retailer running thin margins on physical goods, four points of revenue is most of the profit on the order, which is why the Instant Checkout pilot had a dozen Shopify merchants after five months and not a thousand.

The networks have the opposite problem. They are charging for verification, through programmes, certifications and token issuance, and they are not charging for the liability. When an agent order goes wrong, the merchant pays, the issuer fields the call, and the network keeps its fee.

The asymmetry explains why the protocols are proliferating. Visa’s Trusted Agent Protocol, Mastercard’s Verifiable Intent, Google’s AP2, Stripe’s shared payment tokens and OpenAI’s own checkout specification all describe the same transaction from different angles. A merchant that wants to accept agent orders in 2027 may need to support four of them.

Europe has a rule already, and it points at the human

Europe starts from a different place, because it already regulates who must authenticate. The second Payment Services Directive requires strong customer authentication of the payer. When an agent acts for a payer, the regulation still looks through the agent to the person behind it. Visa’s European pilots handled that with payment passkeys bound to the user, which is why Intelligent Commerce could go live with 30 issuers before the US had a consumer product.

That framing is useful and incomplete. Strong authentication proves the human approved the transaction. It does not define who is responsible when the agent approves something the human did not intend. The directive was written for third-party payment providers: companies with a licence, capital and a supervisor. An agent has none of those things, and no rule published through 2026 assigns it a legal personality.

The gap will be tested first in categories where liability chains are already explicit. Insurance, banking and healthcare all have processes for who pays when a sale goes wrong, and all three are seeing rapid agent adoption. Nearly seven in ten American shoppers already use AI somewhere in the buying process, and about 30% have completed a purchase through it. Those numbers describe a consumer habit that is arriving faster than the rules that govern it.

One rule change would matter more than any protocol

Three developments would move agent payments from a controlled experiment to a real channel, and each is observable.

  • A network publishes a dispute reason code for agent-initiated purchases, or a liability shift attached to a certified agent programme. That single rule change would do more than any protocol release.
  • A large merchant group adopts delayed capture as standard for agent orders, which would turn the notification step into a consumer habit rather than a fraud control.
  • A regulator states whether an AI agent acting for a consumer is a payment initiation service under PSD2 and its successors. The directive was written for licensed entities, and an agent holds no licence.

Until one of those lands, the wallet is a well engineered way to move a purchase decision into a chat window. It will not change how people pay until a network moves the loss. Everything shipped in 2026, including this wallet, is preparation for that decision, with the risk still sitting where it has always sat: on the merchant’s books, in a dispute queue that has no special code for what just happened.

Methodology

Data as of November 2026. Launch details come from OpenAI’s announcement as reported by TechCrunch and CNBC on 13 November 2026, including the card networks involved, the token design and the absence of a transaction fee. Market size figures are taken from the published forecasts named in the table, each of which defines agentic commerce differently; they are not directly comparable and are presented side by side rather than averaged. The $823 million figure multiplies EMARKETER’s 2026 estimate by the 4% fee OpenAI charged on Instant Checkout in 2025, which is an illustration of the revenue at stake, not a forecast of what any firm will earn. Consumer trust percentages come from surveys commissioned by Visa and Accenture, with sample sizes of 2,065 and 25,590 respectively. Dispute rules were checked against network programme documentation available in November 2026 and may change without notice.

Sources

  • TechCrunch 🔗, ChatGPT’s new Wallet lets AI agents shop and check out for you, 13 November 2026
  • CNBC 🔗, OpenAI launches ChatGPT Wallet, 13 November 2026
  • TechCrunch 🔗, United Airlines blocks AI agents from booking flights, 12 November 2026
  • Retention Edge 🔗, AI in ecommerce statistics for 2026, accessed November 2026
  • Payment Review 🔗, a year of agentic checkout, November 2026

About Global Infinity Labs

Global Infinity Labs is an AI innovation company specializing in MCP data services. This article was prepared by our expert team using insights and data provided through our MCP platforms.

Contact: support@globalinfinitylabs.com