Skip to content

Guide8 min readPublished

AI agent fraud and chargebacks: is it safe?

What can go wrong when an AI agent pays, from card testing to “I didn’t buy that”, and what Visa, Mastercard, AP2 and Stripe give a merchant as evidence. No fraud rate for agent orders is published, so we give none.

What happens when an AI agent pays for something and the cardholder says they never ordered it? The same thing that happens with any other card payment, to begin with. Stripe's definition does not change for agents: "A dispute (also known as a chargeback) occurs when a cardholder questions your payment with their card issuer", and the issuer's dispute "immediately reverses the payment."

An agent's card payment falls under the same dispute rules as any other card payment, with one new question attached: can you show that this agent was who it claimed to be, and that the shopper approved this order? The card networks, AP2 and Stripe have each published a way to answer it, and some are still early: as of November 2026, Stripe labels its shared payment tokens "Public preview" and AP2's specification is at v0.2. None of them publishes a fraud or chargeback rate for agent orders, so nobody can honestly tell you whether agent orders are riskier than other orders yet. We will not guess either.

It helps to separate the risks first, because each one has a different answer.

What can go wrong when an AI agent buys from your store?

Four different things, which are easy to lump together as "bots":

Four things that can go wrong with an agent’s order, and what speaks to each
What goes wrong

Showing: A bot testing stolen cards

What it is

Stripe: “the practice of testing a card (or multiple cards) on one site to see if it’s still valid before using it on another site to make a fraudulent payment.”

What the published mechanisms offer

Stripe’s advice: “Implementing CAPTCHA or rate-limiting charges can help combat this type of fraud.” Visa’s Trusted Agent Protocol aims “to help merchants differentiate legitimate agents from nefarious bots or web crawlers”.

Our reading

Rate limits catch the card tester without stopping a shopper’s agent. A CAPTCHA in front of payment stops both.

Sources: Stripe Docs: Common types of online fraud, Visa Developer: Trusted Agent Protocol, getting started and Storewalk: TRX-002, no CAPTCHA before the payment step

Card testing and friendly fraud are old problems. Card testing is a bot that runs through stolen card numbers, and it did not wait for AI to be invented. Friendly fraud, in Stripe's words, "occurs when a legitimate cardholder makes a purchase, but then disputes it at a later date", either because they "didn't recognize the transaction on their statement" or on purpose.

The other two are new. One is an agent that claims to shop for someone, with nothing to back the claim. The other is an agent that the shopper really did send, which then bought the wrong size. Stripe's existing advice for card testing, "Implementing CAPTCHA or rate-limiting charges can help combat this type of fraud", is half helpful here: a rate limit slows the card tester and leaves a shopper's agent alone, while a CAPTCHA at checkout blocks both.

How can a store tell a real shopping agent from a bot?

The signature tells you, where the agent sends one. Visa's Trusted Agent Protocol puts it simply: "a trusted agent will include a message signature in its request to your server. A merchant's primary task is to validate this signature." Visa presents it as "a verifiable, cryptographic standard to help merchants differentiate legitimate agents from nefarious bots or web crawlers." Mastercard's framework rests on "registering and verifying AI agents before they are permitted to transact on the Mastercard network", and suggests checking signatures with Web Bot Auth "at the Content Delivery Network (CDN) layer" so that "merchants can verify agent authenticity without deploying new code".

Here is what a signature header looks like, and what Visa's rules check in it. Try the samples, or paste one from your logs:

Check a Trusted Agent Protocol Signature-Input header against Visa’s merchant rules
  1. Tag is agent-browser-auth or agent-payer-auth: passestag="agent-browser-auth"
  2. Minimum fields present: passesAll eight are there
  3. created and expires at most 8 minutes apart: fails3600 seconds apart
  4. created in the past, expires in the future: not runNeeds a clock: the page runs this step against yours

Visa: “the message should be blocked.”

What this page cannot check: a nonce seen in the last 8 minutes, the agent’s public key, and the signature over the request. Visa’s spec blocks the message on any of those too. Nothing you paste leaves the page.

Source: Visa Developer: Trusted Agent Protocol specifications

A valid signature tells you which registered agent sent a request. It does not tell you that the shopper wanted this particular order. That is the job of the next layer, and the one that matters most for chargebacks. Bots, crawlers and agents covers Web Bot Auth itself in more depth.

What evidence do you have if an agent's order is disputed?

With a person at your checkout, your evidence is whatever you logged. With an agent on these programmes, there can be a signed record of what the shopper approved before anything was bought. Pick a mechanism:

What each agent payment mechanism gives a merchant, in its owner’s words
The mechanism

Showing: Visa Trusted Agent Protocol

What it gives you

“a verifiable, cryptographic standard to help merchants differentiate legitimate agents from nefarious bots or web crawlers.” It also “describes signatures to enable consumer recognition and payment.”

Who holds it (our reading)

The agent signs each request; you, or your CDN, check the signature.

What you do

“A merchant’s primary task is to validate this signature.”

Source: Visa Developer: Trusted Agent Protocol, getting started

Under each name is the owner’s own status label, version or date, read on 4 November 2026; “no status label” means the page has none. None of these pages states a fraud or chargeback rate for agent orders, and we do not estimate one.

The wording is strong in places. AP2's specification says its checkout and payment mandates and receipts "can be brought together to provide a non-repudiable picture of the transaction." Visa's purchase instructions are "Required to evidence the specific authorized transaction for disputes". Mastercard calls its purchase intent data "an audit trail that may be used to help avoid and/or resolve potential cardholder disputes."

Read the small print too. AP2 says "Specific details of how this is used for dispute resolution, retention, and retrieval requirements are outside the scope of this specification." Mastercard's sentence says "may be used". None of the pages we read says that a chargeback on an agent order will be decided differently because the evidence exists. The evidence can be there and still not be in your hands when the dispute arrives, so ask your payment provider where it goes. AP2 explained and Visa Intelligent Commerce vs Mastercard Agent Pay go through each programme step by step.

Do agent payment tokens reduce fraud?

They narrow what a stolen or misused credential can do, which is what a merchant wants from them. Stripe's shared payment token "is a scoped grant to use the customer's payment method", and the agent grants it "each with usage and expiration limits". Stripe adds that "The agent sets the maximum amount to match the total amount of the transaction" and that "You can't create a payment with a revoked SPT." As of November 2026, shared payment tokens are labelled "Public preview". Mastercard describes its agentic tokens as credentials "that ensure every transaction is both traceable and authenticated."

For you, a token like this means the agent never held the card number, and the charge cannot be larger than the limit the shopper's side set. Whether that lowers disputes on your store is something no publisher we read has measured. How agents pay explains the token types.

Who is responsible for an agent's mistake?

None of our sources says. If an agent buys the wrong size, the shopper did ask for a purchase, so it is not fraud in Stripe's sense, and none of the protocols assigns the cost of an agent's error. To us, that makes it a return under your own policy. OpenAI's commerce docs, for one, put risk on the merchant's side of the table: the merchant "analyzes payment and risk signals on their own stack". The merchant of record article covers who handles refunds on each platform.

Write a return policy an agent can read, keep order confirmations clear about what was bought, and make your store's name on the card statement recognisable, since an unrecognised charge is the accidental half of friendly fraud. The shipping and returns guide helps with the first.

How do you protect your checkout without blocking good agents?

Put the protection where it catches fraud and lets shoppers through. Rate limits and bot management catch card testers. A CAPTCHA in front of the payment step stops any agent that cannot hand it to a person, the honest ones included, and leaves you with the fraud filters you already had. The bot protection guide walks through vendor settings.

Our scan checks the parts it can see from outside. TRX-002 fails a CAPTCHA on the cart or checkout. VER-001 looks for bot management in front of the store, and VER-003 compares the controls on your checkout with those on a product page. It does not see your fraud filters, your disputes or your payment provider's settings. A CAPTCHA is only one of the ways a store turns agents away by accident. Whether yours is one of them is a question TRX-002 answers when you run our agent readiness checker.

Frequently asked questions

Are AI agent purchases more likely to be fraud?

No one we read has published data on it. Visa, Mastercard, AP2 and Stripe describe ways to verify agents and record what shoppers approved, but none gives a fraud or chargeback rate for agent orders.

Can a shopper charge back a purchase an AI agent made?

Yes, as with any card payment. A dispute "occurs when a cardholder questions your payment with their card issuer", whoever clicked the button. The difference is the evidence you may have, such as an AP2 mandate or Visa purchase instructions.

What is friendly fraud?

Stripe's term for a legitimate cardholder who makes a purchase and disputes it later, by accident because they did not recognise the charge, or on purpose. An agent order the shopper forgot approving could turn into one.

How do I know a request really comes from a trusted AI agent?

Check its signature. Under Visa's Trusted Agent Protocol and Mastercard's framework, registered agents sign their requests, and the merchant or its CDN validates the signature with the agent's public key.

Should I block AI agents to prevent fraud?

Blocking all of them stops sales as well as fraud. Rate limits and signature checks separate card testers from shopping agents better than a CAPTCHA at checkout, which stops both.

The sources below were last checked on the date shown.