Agentic Payments: How Agents Pay, and Who's Liable When They're Wrong

Agentic payments explained: how agents get tokenized cards, spending mandates and approvals, 7 rails compared, and who pays for a wrong buy. Read on.

Quick Answer: Agentic payments are payments an AI agent makes with a tokenized credential tied to that agent, inside a spending mandate you set, with human approval above a threshold. The rails exist as of November 2026; the liability rules don't. A hijacked agent looks like fraud, but an agent that buys the wrong thing within its limits falls into an open gap.


Agentic payments are the moment an AI agent stops recommending and starts spending. The agent holds a credential, decides a purchase fits what you asked for, and pays, sometimes while you're asleep. Visa, Mastercard, Google, OpenAI and Stripe, Coinbase, PayPal and Adyen have all shipped or announced rails for it.

Coverage of how agentic commerce is reshaping retail focuses on shopping and the protocol race. This explainer covers the plumbing underneath: how an agent actually gets permission to pay, what each rail does as of November 2026, and the question most write-ups skip. Who pays when the agent is wrong?

We build cowork.ink, where AI coworkers do research and prep work in their own browser and hand decisions back to people. The checkout button is where an agent should hesitate most.


What are agentic payments?

An agentic payment is a transaction an AI agent initiates and completes on someone's behalf, using delegated authority with limits attached. What changes is who presses the button, and what the network has to check before it says yes.

A normal card payment verifies one thing: that the cardholder can pay. An agentic payment has to verify three:

  1. The agent is who it says it is, and not a scraper or a bot pretending to be one.
  2. The person delegated this purchase, at this amount, at this merchant.
  3. The agent stayed inside that delegation when it built the cart.

That is also the difference from autopay. A subscription renewal follows a fixed rule. An agent picks the product, the merchant and often the timing, so the system needs proof of what it was allowed to choose.


How does an AI agent actually pay?

Through three controls stacked on top of each other: a tokenized credential, a spending mandate, and a step-up approval for anything outside it.

Tokenized agent credentials

On the dedicated rails, the agent never sees your card number. The card is swapped for a token that only works for that agent, often only at one merchant or up to one amount. (A browser agent filling in a checkout form with your saved card is a different and riskier setup.)

On Visa Intelligent Commerce, the card is enrolled and provisioned as a token, and the agent stores only the token ID. On Stripe, the agent issues a Shared Payment Token scoped to one seller, with a currency, a maximum amount and an expiry, and it can revoke that token at any time. Stripe's documentation notes it may use Mastercard Agent Pay and Visa Intelligent Commerce network tokens underneath.

A token tied to one agent answers "which agent did this?", the question behind our guide to AI agent identity.

Spending mandates and limits

A mandate is the written-down version of what you asked for, signed so nobody can edit it afterwards. Google's Agent Payments Protocol (AP2) is the clearest example. Per the AP2 specification, a user's constraints start as an "open" mandate, such as a budget or a product description, and become a "closed" mandate once a specific checkout and amount are final. Typical limits are an amount cap, a merchant allowlist, an expiry and whether nobody needs to be present.

Step-up human approval

When a purchase falls outside the mandate, the agent should stop and ask. Visa's developer documentation describes passkeys for cardholder consent and step-up authentication through a one-time code or a banking app. Stripe's tokens can enter a requires_action state when a bank asks for 3D Secure, which hands control back to the person.

Our guide to human-in-the-loop AI agents covers where to set approval points people won't learn to click through.


The agentic payment rails, as of November 2026

Seven rails matter right now, each solving a different layer. Card networks handle agent identity and tokens, open protocols carry proof of what the user approved, and processors connect both to merchants. Status is what each sponsor's own announcement or documentation states.

RailSponsorHow authorization worksStatus (Nov 2026)
Visa Intelligent CommerceVisaAgent token; passkeys confirm the cardholder; consumer-approved purchase instructions; Trusted Agent Protocol identifies verified agentsDeveloper program (sandbox to production). Intelligent Commerce Connect, one merchant integration, announced in pilot April 2026
Mastercard Agent PayMastercardAgentic tokens; agents identify with Web Bot Auth; Verifiable Intent records what the user approvedAnnounced April 2025. Verifiable Intent is an open draft spec, contributed to the FIDO Alliance in April 2026
Agent Payments Protocol (AP2)Google, now FIDO AllianceSigned Checkout and Payment Mandates, from open constraints to a closed final amountv0.2, human-present and not-present flows. Moved to the FIDO Alliance, April 2026
Agentic Commerce Protocol (ACP)OpenAI and StripeShared Payment Token scoped to one seller, amount and expiry; merchant stays merchant of recordFirst implemented in ChatGPT. Token API in preview (US, Canada, parts of Europe)
x402Created by Coinbase; x402 FoundationServer answers HTTP 402; the client pays, usually in stablecoins, and retriesUnder the x402 Foundation at the Linux Foundation; x402.org calls it production-ready
PayPal agent readyPayPalPayPal merchants accept payments on AI surfaces, with PayPal's buyer protection and dispute resolutionAnnounced October 2025 for early 2026
Adyen AgenticAdyenAuthentication, token portability and risk across AP2, ACP and UCPAnnounced June 16, 2026; limited availability, US enterprise

Two patterns stand out. First, the networks have converged on agent identity plus tokens: Cloudflare announced in October 2025 that Visa, Mastercard and American Express would use its Web Bot Auth so merchants can recognize trusted agents. Second, the protocols have converged on signed proof of intent, and that proof is exactly what a dispute will turn on.

Our map of the AI agent protocol stack shows where these sit next to MCP and A2A.


Who is liable when an agent buys the wrong thing?

As of November 2026, we found no law or network rule that says "if an AI agent made the purchase, this party pays." Liability falls out of older frameworks written for people, and the answer depends on which kind of wrong it was.

General information, not legal advice

This section describes how existing US rules and card-network processes are written. It is not legal advice, and none of these rules has been tested against agent purchases in a published regulator decision we could find. For a real dispute or a product decision, talk to a payments lawyer.

Authorization is not intent

Authorization means the agent was allowed to pay. Intent means the purchase is what you actually wanted. Payment law is built around the first. Agent mistakes live in the second.

If someone steals your card, the purchase was never authorized, and decades of rules protect you. If you hand your agent a token with a $200 limit and it buys the wrong $180 jacket, the payment was authorized. It just wasn't what you meant.

What US Regulation E and Regulation Z say

The two main US consumer protections were written with "a person with your card" in mind, and their text reads awkwardly against an agent.

  • Regulation E (debit and bank transfers). It defines an unauthorized electronic fund transfer as one "initiated by a person other than the consumer without actual authority." It then excludes transfers by "a person who was furnished the access device to the consumer's account by the consumer," unless the consumer has told the bank those transfers are no longer authorized. Read plainly, a credential you gave your agent sits closer to the exception than to the protection.
  • Regulation Z (credit cards). Unauthorized use means use by someone without "actual, implied, or apparent authority," and the cardholder's liability for truly unauthorized use is capped at the lesser of $50 or the amount charged before the issuer is told.

Neither rule mentions agents, and whether an agent that ignored its mandate had "actual authority" is still an open question.

How card disputes fit

Card disputes fall into broad buckets; Visa's process uses four: fraud, authorization, processing errors and consumer disputes. An agent's honest mistake fits none cleanly. It isn't fraud, the authorization was valid, and the merchant processed and delivered what was ordered.

What went wrongClosest existing frameWhat is still unsettled
Attacker steals the agent's token or accountClassic unauthorized use: fraud dispute, Reg E or Reg Z protectionsWhether a token you handed over counts as furnished to the attacker
Agent buys the wrong item, inside its mandateAuthorized purchase: merchant return policy or consumer disputeWhether 'my agent misread me' is a valid dispute reason
Agent exceeds its mandatePossibly unauthorized, if the mandate is provableWhether exceeding a mandate removes actual authority
Prompt-injected purchase at an attacker's shopFraud by a third party, but paid with a valid credentialWhether the agent provider shares the loss
Merchant ships something other than the cartOrdinary consumer dispute against the merchantLittle: the signed cart helps the buyer here

Where it is heading

The industry is answering with evidence first and guarantees second. AP2 describes its mandates as "a non-repudiable, cryptographic audit trail for every transaction." Mastercard says Verifiable Intent links identity, intent and action into one record that all parties can rely on in a dispute. Expect the party with the better record of what the user approved to have the stronger case.

The first explicit guarantee came from American Express. In April 2026 it said that "in the future" it will protect eligible card members from charges caused by errors of registered AI agents, when the member authorized the agent and the agent sent Amex an authenticated purchase intent. No launch date was given.

The practical takeaway: an AI agent audit trail is no longer just an engineering nicety. In agentic payments, it is your evidence.


What are the fraud risks of agentic payments?

Two risks are new: real agents talked into buying, and fake agents that talk merchants into trusting them.

Prompt-injected purchases

An agent that reads web pages can read instructions planted in them. A product listing, an email or a review can tell it to add an item, change the shipping address or pay a different merchant. The payment then goes through with a perfectly valid credential, so nothing in the logs looks like an attack.

OpenAI said in December 2025, while hardening its Atlas browser agent, that prompt injection is unlikely ever to be fully "solved." So the defense is structural, not clever prompting: tight mandates, merchant allowlists, and a human approval step for anything new.

Agent impersonation

A scraper or fraud bot can claim to be a well-known shopping agent. That is why Visa's Trusted Agent Protocol and Mastercard's use of Web Bot Auth have agents sign their requests, so merchants check a signature instead of trusting a user-agent string.


What should businesses do about agentic payments now?

Treat every agent that can spend like a new employee with a company card: named, limited, logged and easy to cut off. The specifics depend on which side of the checkout you're on.

If your team uses agents that can pay:

  1. Give each agent its own credential. One token per agent, never a shared corporate card number pasted into a prompt.
  2. Write the mandate down. Amount caps, merchant allowlists and an expiry, narrower than feels necessary.
  3. Set an approval threshold. Anything above it, or at a merchant the agent hasn't used before, waits for a person.
  4. Keep the record. Save the mandate, cart and approval for every purchase.
  5. Review the guardrails in front of the payment step, especially against untrusted web content.

If you sell to agents:

  1. Ask your processor which agent rails it supports. Visa's Intelligent Commerce Connect, for one, is pitched as a single integration covering several protocols.
  2. Recognize signed agents instead of blocking all automated traffic.
  3. Keep clear return terms. A fast return beats a dispute for everyone.

Let a Coworker Do the Legwork, Keep the Card

Most of the value in agentic purchasing comes before the payment: comparing vendors, reading terms, shortlisting options. An agent can do that part today with no liability question at all.

At cowork.ink, each AI coworker has its own computer with a real browser, files and an inbox. Hand it the research: compare three suppliers, pull their pricing pages, read the cancellation terms and post a shortlist with sources in Slack or Telegram. Every run leaves a transcript of the pages it opened, so you can check its work. Then you make the decision and you press pay.

Setting one up takes about three minutes. It's $25/month billed yearly, or $29 monthly, for every coworker you set up, with a 15-day money-back guarantee.

Frequently Asked Questions

What are agentic payments?
Agentic payments are payments an AI agent starts and completes for a person or a business, using a credential it was given and inside limits that person set. Instead of only checking that the cardholder can pay, the system also checks that this agent is allowed to pay, for this amount, at this merchant.
How do AI agents make payments?
On the agent payment rails, the agent never holds your card number. It gets a token tied to that agent (a Visa or Mastercard network token, or a Stripe Shared Payment Token), a mandate that sets the limits, and a rule for when it must stop and ask you to approve. Our guide to [AI agent permissions](/blog/ai-agent-permissions/) covers how to scope those limits.
Who is liable if an AI agent makes a wrong purchase?
As of November 2026 there is no single rule. Stolen or hijacked credentials look like classic fraud, which consumer protection law covers. A purchase the agent was allowed to make but got wrong is closer to a merchant dispute or a return, and the signed record of what you approved is likely to be the key evidence.
What is the Agent Payments Protocol (AP2)?
AP2 is an open protocol Google released in 2025 and handed to the FIDO Alliance in April 2026. It records what a user approved as signed mandates, one for the checkout and one for the payment, so merchants, networks and issuers can check the agent stayed inside what the user allowed.
Are agentic payments safe?
The credential side is reasonably safe because agents get scoped, revocable tokens instead of card numbers. The weak point is the agent's judgement: prompt injection can steer an agent to a purchase you never wanted, which is why spending caps and a human approval step still matter.
Home Blog Company