top of page

JPMorgan Blocked the Agents. Stripe Priced Them. The Agentic Acceptance Fork Is Now Live.

Nov 14
7 min read

TL;DR

The agentic payments debate has moved beyond whether banks let AI agents access accounts. The new question is what happens at the merchant’s terminal when an agent pays. JPMorgan is blocking agents from shopping with its customers’ card numbers, while Stripe Link can route an approved purchase through a single-use virtual card. But merchants say those cards are classified as commercial cards, which can cost more to accept and complicate returns. For payments operators, the agent’s acceptance route is now an economics and operations decision, not just a question of access.

What did JPMorgan actually block?

JPMorgan Chase is blocking AI agents from shopping with its customers’ credit cards when a customer gives the agent their card number, according to the Wall Street Journal, as reported by PYMNTS. Cards from Bank of America, Capital One and Citigroup are already being used by agents when customers allow it. The issuer landscape is splitting: some cardholders can use their cards directly with agents, while JPMorgan customers cannot.

This follows JPMorgan’s earlier restrictions on outside agents logging in with customer credentials, sending money or paying bills. After its 13 October earnings call, CEO Jamie Dimon explained the caution: “You don't want rogue agents, and there could be thousands of them.”

The decision matters because the bank’s card book is substantial. JPMorgan’s debit and credit card sales volume reached $539.5 billion in Q3 2026, up 10% from $492.3 billion a year earlier. Yet an issuer block does not necessarily stop a customer from using an agent to spend their own money. As Cardless president Michael Spelfogel told the Journal, banks “can't stop people from using agents to spend their own money.”

Several banks plan to treat a cardholder who gives an agent a card as they would someone who hands a card to a family member: the cardholder remains responsible for the charges, the Journal reported. That is distinct from the bank-side account-access and consumer-liability debate explored in our Chime and Wells Fargo analysis.

Why is an agent's card more expensive for a merchant to accept?

A single-use virtual card can get an agent through checkout, but it can change the economics of the transaction. With Stripe Link, the agent receives a virtual card for one purchase and one approved amount. It pays the merchant with that card, then Stripe bills the shopper’s own card. The merchant is paid by a different card from the one the shopper expected to use.

That distinction shows up in classification. Merchants have told Visa and Mastercard that these virtual cards cost more on average to accept because they are classified as commercial cards rather than consumer cards, the Journal reported. Merchant executives also said temporary cards have made returns harder to process. The shopper sees a convenient agent purchase; the merchant may see a different card category, a different acceptance cost and a less straightforward refund path.

An abstract merchant payment ledger shows a transaction classified differently from the shopper’s expectation

Link has more than 300 million users and is accepted at more than 1 million businesses, according to PYMNTS. At those businesses, an agent such as Meta’s Muse can use the shopper’s saved payment method. Elsewhere, Link issues the single-use virtual card. Anthropic’s Claude Duet buys business supplies in a similar way.

The alternative is an agentic token from Visa or Mastercard. These tokens hide the card number and carry the shopper’s spending limits. In principle, a tokenised transaction can avoid substituting a commercial virtual card for the shopper’s consumer card, preserving the intended consumer-card economics where the issuer and implementation support it. JPMorgan has been slow to adopt these tokens, people familiar with the matter told the Journal. Mastercard CEO Michael Miebach has called a virtual card number “a bridge to get ready for a fully tokenized transaction.”

Issuer-side products are also emerging. Cardless launched Cardless AI on 13 November, combining conversational card servicing and virtual cards that cardholders can give an agent with limits they set. Its Parallax model detects AI-generated or altered images and synthetic identities; the company said about one-third of fraud rings identified in its beta had activity across multiple identity-verification providers. This is one response to the broader challenge of controlling agent spending without turning every agent purchase into an opaque merchant-side event. For related merchant controls, see our analysis of Walmart’s agent-shopping control plane.

Why did United say no while Travelport and Mastercard said yes?

United’s decision is about access to its digital sales channel; the Travelport, Booked AI, Basis Theory and Mastercard demonstration is about designing an agent transaction that keeps the airline in control.

United is restricting AI agents’ access to its digital platforms, directly and through travel agencies and aggregators, and plans to update its booking, ticketing and distribution policies. Its stated concerns include exposing sensitive information when customers share credentials, unintended purchases or account changes, unauthorised use of frequent-flyer miles, and regulatory exposure where a human acknowledgement is required, such as a check-in confirmation about prohibited items. Reuters reported the move on 11 November.

At the same time, Expedia and Booking are working with OpenAI, while Amazon has blocked Meta’s Muse and Walmart has partnered with it. The market is testing different approaches to who controls the agent’s route into a merchant’s platform.

United’s position is not the only possible design. In the live flight booking announced on 12 November, an agent used existing travel and payment infrastructure, with the airline remaining the merchant throughout. The contrast is not simply “allow agents” versus “block agents”. It is whether the channel can be designed so access, payment and customer approval remain controlled.

What does the live flight booking actually prove?

It proves that an agent can search, book, pay for and ticket a flight within a conversation without seeing the underlying card number, while the airline remains the merchant. It does not prove that this flow is already a high-volume business.

In the Travelport announcement, Booked AI’s agent sourced a live NDC offer through Travelport TripServices. The customer confirmed the flight and fare in the conversation; Mastercard Agent Pay used a tokenised credential, and the underlying card details stayed in Basis Theory’s PCI-compliant vault. The airline’s and Travelport’s systems processed and ticketed the booking after authorisation.

Mastercard’s Verifiable Intent linked the customer’s approval of the specific booking and amount to the payment. If the transaction had not matched that approved intent, it would not have proceeded; another purchase would have required new cardholder approval.

A tokenised credential and approval signal travel through connected airline and payment infrastructure

This is a meaningful acceptance design, but it remains one live transaction, not evidence of a scaled volume business. Structured, higher-value categories such as travel may provide an early proving ground. Everyday checkout still has to resolve the merchant’s acceptance cost, classification and returns workflow. For earlier agent payment activity, see our coverage of Visa’s live trusted-agent protocol pilot.

What should payments operators do now?

Payments operators should test the full merchant-side economics and operating path before agent transactions arrive at scale.

First, classify agent traffic on the statement. Determine whether the merchant can identify agent-originated transactions at all, and whether it can distinguish consumer-card tokens from commercial-classified virtual cards.

Second, model the cost difference. Test how a commercial-classified virtual card moves the merchant between pricing buckets, and decide up front who absorbs the gap.

Third, design returns and disputes for single-use cards. Support teams need to match a refund to the original transaction and produce evidence when the payer was an agent. Our article on multi-merchant evidence for card-absent fraud covers the related dispute questions.

Fourth, set an explicit channel policy. Decide whether agents can reach merchant channels directly, through approved partners, or not at all, and review it the way United is reviewing its booking and distribution terms.

Fifth, treat tokenised credentials as the destination and virtual cards as the bridge. The bridge works, but the merchant should know what it costs to cross it.

There is also a direct Australian consideration. With surcharging banned here, a commercial-card cost gap on agent transactions cannot be passed on as a surcharge. It comes straight out of the merchant’s margin. That makes classification an acceptance issue with immediate commercial consequences, not a back-office detail. The broader interchange debate is already contested, as our coverage of merchants challenging interchange-law pre-emption explains.

Consumer expectations point to a measured approach. In a PYMNTS Intelligence study of 5,241 consumers in the US, Brazil and UAE, 52% said they want to choose how to pay when an agent lines up a $300 purchase; another 25% would accept an AI recommendation but retain final approval. Acquirer readiness is also uneven: 23% plan to accelerate client readiness, while 77% will move when clients, platforms or consumers show demand.

What happens next?

Watch whether more issuers follow JPMorgan or allow direct agent use as Bank of America, Capital One and Citigroup reportedly do. Also watch whether Visa and Mastercard change how agent-issued virtual cards are classified, whether merchants refuse agent traffic on cost grounds, and whether travel becomes the proving ground for controlled acceptance.

Distribution terms will matter too. As merchants decide who can reach their checkout and under what conditions, platform policies and payment acceptance rules will be written together. The related Sardine analysis of false declines examines a separate but connected merchant concern: agent sessions that can look like fraud.

FAQ

Why is JPMorgan blocking AI agents from using its cards?

JPMorgan is blocking agents when customers give them their card numbers. The bank had already restricted outside agents from accessing customer accounts, sending money or paying bills. CEO Jamie Dimon cited the risk of “rogue agents”.

Why might a virtual card cost merchants more?

Merchants told the Journal that the single-use virtual cards used for some agent purchases are classified as commercial cards rather than consumer cards, making them more expensive to accept on average.

Does Stripe Link always use a virtual card for agent purchases?

No. At businesses that accept Link, agents can use a shopper’s saved payment method. At other businesses, Link issues a single-use virtual card for the approved purchase.

What did the Travelport and Mastercard flight booking demonstrate?

It showed one end-to-end flight booking and payment in which an agent used a tokenised credential, the customer’s approval was linked to the transaction, and the airline remained the merchant. It was a live transaction, not evidence of a scaled business.

What should Australian merchants check first?

They should identify how agent transactions are classified, model the acceptance cost and confirm how returns and disputes will work. Because surcharging is banned, a higher acceptance cost cannot simply be passed on as a surcharge.

Key takeaway: Agentic commerce is now an acceptance-economics question. A transaction can be approved by the shopper and still arrive at the merchant as a different card category, with different costs and operational consequences. Payments operators should measure that gap now, build a workable returns path and decide which agent channels they are prepared to serve.

 
 
bottom of page