For a cardholder, the Visa Trusted Agent Protocol and Mastercard Agent Pay mostly change what happens behind the checkout, not what you do at it. An AI agent pays with a token tied to that agent instead of your card number, you approve the setup and the purchase, and the networks pass extra “this was an agent” data to merchants and issuers. Your bank does not have to ask you to sign up.

How we checked this

We read Visa’s and Mastercard’s own releases, stories and developer pages on Visa Intelligent Commerce, the Trusted Agent Protocol, Agent Pay and Agentic Tokens, published between April and December 2025. For your rights as a cardholder, we used the CFPB’s page on disputing a credit card charge. Facts checked on September 24, 2026.

What the two programs are, in one table

Both networks built on something that already exists: network tokenization, the same technique that puts a stand-in number on your phone for contactless payments. The difference is that the token is issued for a specific AI agent, and the transaction carries extra data about the agent and what you asked it to do. For the protocol-level picture, our agentic commerce protocol tracker lists who supports what.

Visa Mastercard
Program names Visa Intelligent Commerce; Trusted Agent Protocol (TAP) Agent Pay; Agentic Tokens
Launched Intelligent Commerce April 30, 2025; TAP October 14, 2025 Agent Pay April 29, 2025
Credential used Agent-specific tokens provisioned by Visa Agentic Tokens built on Mastercard tokenization
How the agent is recognized Cryptographic agent signatures (HTTP Message Signatures) Agent registration plus Web Bot Auth at the CDN layer
Your approval Step-up verification and a passkey at setup; agent asks you before each purchase “Strong consumer authentication” with on-device biometrics; permissions and limits you set
Data for disputes “Commerce signals” plus your instructions Purchase intent data and an audit trail
U.S. cardholder availability Announced for all U.S. cardholders by mid-November 2025

Visa’s April 2025 release describes “AI-ready cards” that use tokenized credentials instead of card details. The Trusted Agent Protocol release is aimed at merchants: it lets them tell an approved agent from a bot using agent-specific cryptographic signatures, and says the agent can optionally carry payment data. Mastercard’s April 2025 launch says agents must be “registered and verified” before they pay and that Agentic Tokens build on the tokenization used for contactless and card-on-file payments.

What an agent-scoped token means for you

A normal card-on-file payment hands the merchant (or the agent) your card number and expiry date. If that data leaks, anyone can try to use it anywhere. An agent-scoped token is different in two ways that matter to a cardholder.

First, the token is not your card number. Visa’s developer page for Visa Intelligent Commerce says the system provisions tokens meant for agent transactions and handles their “provisioning and life cycle management.” Mastercard’s September 30, 2025 story calls Agentic Tokens “dynamic digital credentials” and says each transaction is “tied to a specific, authorized Mastercard Agent Pay interaction so that you understand which agent acted on your behalf.”

Second, the token carries limits. Visa’s April release says consumers can “set spending limits and conditions.” Mastercard’s framework story from October 14, 2025 lists “cart contents, transaction limits and validity windows” among the purchase intent data. Visa’s developer page adds that controls check “that the request originates from the intended merchant for the correct amount.” Payments processor Checkout.com, in its January 2026 trends piece, summarized the change the same way: Visa updated its network token provisioning to give AI agents “context-specific payment credential use.”

Where those limits are set, and whether you can edit them after setup, depends on the agent and your bank. Our spending limits guide walks through what ChatGPT, Gemini and card issuers let you control today.

Does my bank have to support it?

Not in any way you need to act on. Both networks frame the programs as network features rather than something a cardholder applies for, and Visa’s developer page describes issuers controlling token activation and termination in the background.

Mastercard was the most specific. Its September 10, 2025 release said all U.S. Mastercard cardholders would be enabled for Agent Pay by the 2025 holiday season, with Citi and U.S. Bank cardholders first. The September 30 story said Citi and U.S. Bank cardholders were already enabled, all U.S. cardholders would follow by mid-November 2025, and global cardholders “shortly after.” It also said OpenAI’s Instant Checkout was the first place Agent Pay was used.

Visa’s December 18, 2025 update reported “hundreds” of completed agent-initiated transactions, more than 100 partners, and pilots with Skyfire, Nekuda, PayOS and Ramp. It predicted that “millions of consumers” would use agents to buy by the 2026 holiday season. That is a forecast, not a measure of how many Visa cards can be used this way today.

The early access was described by Mastercard, not by the banks. In practice, “enabled” appears to mean your card can be tokenized for an agent when you add it inside that agent. If your issuer publishes a page, it will be the authority for your card.

The consent happens inside the agent, not at your bank. Visa’s developer page describes the flow: the agent lets you create an account and add your Visa card, the system runs “step up verification of the cardholder as well as setting up a Passkey,” and when the agent finds something to buy, it “will request whether the user agrees that the agent can make the purchase on the user’s behalf.” Visa’s April release puts it plainly: “Only the consumer can instruct the agent on what to do and when to activate a payment credential.”

Mastercard says consumers have “complete control over what the agent is allowed to purchase,” with authentication “leveraging on-device biometrics.” Its September 2025 release mentions work with the FIDO Alliance on a credential that confirms the amount, merchant and product so the network can show the shopper approved them.

So there are two consent moments: once when you add the card and set up a passkey, and again when a purchase is proposed. How strict the second one is depends on the agent. An agent may ask every time, or it may act within limits you set earlier, which is what Mastercard means by a “recurring payment” done “with the permissions and limits you define.”

Visa mentions life cycle management of tokens, which includes shutting a token down. The practical routes are to remove the card from the agent’s account settings and to ask your issuer to delete the agent token. Visa’s developer page says issuers take part in activating and ending tokens, so the bank is the party that can confirm a token is gone. Ask about the agent token by name rather than assuming a card replacement handles it.

What the issuer sees, and whether your statement changes

Issuers get more context than they do for an ordinary online purchase. Visa’s April release says “commerce signals are shared in real-time with Visa,” and the developer page says issuers receive those signals together with the user’s instructions. Mastercard’s framework lists “trusted agent recognition, purchase intent information, agentic tokens themselves, and consumer identity” as standard data elements, and its April release says the goal is that “every player in the value chain, from consumers to issuers and merchants” can recognize an agent-made transaction.

That means your bank can, in principle, tell that a charge came through an agent, which agent, and roughly what you asked it to do. That helps fraud screening and helps the bank judge a dispute.

What your statement shows is up to the issuer’s statement and app design. Assume the charge will look like a normal purchase from the merchant and keep your own record of what the agent bought.

For what the agent itself can see of your finances, a separate issue from the card network, see what your AI assistant sees when you share bank data.

How disputes work when an agent paid

The networks built these programs partly to make disputes easier to settle. Visa’s developer page says the commerce signals “along with the user instructions can be used to resolve any disputes.” Mastercard’s framework promises “an audit trail that may be used to help avoid and/or resolve potential cardholder disputes.”

Your existing rights still apply. For a U.S. credit card, the CFPB says to call the card company right away and send a written billing error notice within 60 calendar days after the charge appeared on your statement. The company then has 30 days to acknowledge it. If the charge is an error, it must be removed; if not, the issuer must explain why in writing.

The open question is what the new data does to your case. If you told the agent “buy the blue one under $80” and it bought the blue one for $79, the intent record supports the merchant, not you. If the agent bought something outside your instructions, the same record may help you. That is the trade: better evidence in both directions. Our liability guide covers refunds, chargebacks and who pays in more depth. If a large amount is at stake and the issuer refuses, a consumer-protection attorney or your state attorney general’s office can tell you what applies to your card type.

What if I never want an agent to use my card?

The documents say that nothing happens without you. Visa’s developer page describes the agent token being set up when you add your card to an agent and authenticate, Mastercard ties each Agentic Token transaction to an interaction you authorized, and Visa states that only the consumer can tell the agent when to use the credential. If you never add a card to an agent, there should be no agent token to use. The residual risk is the old one: someone who has your card number or account login. That is a fraud problem, not an agent problem, and the usual defenses apply. See our 12-point account protection checklist.

If you want to be sure, you can ask your issuer three questions: whether any agent tokens exist on your card, whether they can be deleted, and whether the bank can block new ones. Answers will vary by bank.

What to watch out for

  • Fake agents. The networks verify registered agents; an app pretending to be an agent is still just an app asking for your card. Our guide to fake AI shopping agents shows how to tell them apart.
  • Standing permissions. A limit set once and forgotten can keep working. Check what limits and validity windows you agreed to at setup.
  • Removing the card is not the same as deleting the token. Ask your issuer about agent tokens specifically.
  • Rollout dates are announcements. Mastercard’s mid-November 2025 date and Visa’s 2026 forecast come from the companies. Whether a specific card is covered is for its issuer to confirm.
  • Statement blind spot. If your statement does not name the agent, the intent record exists but you may not see it. Keep screenshots of what you approved.

Go deeper