Who Pays When an AI Agent's Purchase Is Disputed? Chargebacks and Evidence for Agentic Orders
The Question Every Agentic Checkout Leaves With the Merchant
Getting recommended by an AI agent is the visibility problem. Getting paid, and staying paid, is the next one. Once an agent can complete a purchase inside ChatGPT, Google's AI Mode, Gemini, or Perplexity, a new kind of dispute becomes possible: the shopper says "I didn't buy that, my AI did," or "that's not what I asked for."
The industry has not settled who should absorb that loss. But for most merchants today, the practical answer is already clear from the platforms' own documentation: you do. This post explains why, what has changed in the last few weeks, and the evidence you should start capturing on agent orders now.
Why Disputes Land on the Merchant
The deciding fact is merchant of record. Chargeback liability attaches to the merchant of record, and every major agentic checkout program is designed to keep that role with you.
- Google (UCP-powered checkout). Google's developer and Merchant Center documentation states that merchants remain the merchant of record when checkout happens on Google surfaces.
- OpenAI (Agentic Commerce Protocol). OpenAI's commerce documentation states that OpenAI and its payment providers are not the merchant of record. Its payment guidance also says your business handles the order, fulfillment, customer support, and fraud decisions.
- Stripe (Shared Payment Tokens). Stripe's seller documentation says that once you confirm a payment with a shared payment token, subsequent events such as refunds and reporting behave as if you had received the payment method directly.
In other words, the agent platform brokers the purchase, but the transaction is yours: your refund policy, your fulfillment, your fraud decision, and your dispute ratio with the card networks. Agent-initiated disputes count against you the same way any other card-not-present dispute does.
What Makes Agent Disputes Different
Card disputes were designed around a human clicking "buy." Agent purchases break that assumption in three ways.
- The buyer did not see the final screen. A shopper may have delegated "find me running shoes under a set budget" and never reviewed the specific product, size, or seller.
- "Authorized" and "intended" can diverge. A transaction can be properly authorized and correctly fulfilled and still not match what the shopper meant. Chargebacks911's CEO has publicly argued that existing chargeback reason codes were not built for that scenario.
- Fraud and friendly fraud blur. Real agent mistakes, a customer changing their mind, and bad actors using automation all show up as "I didn't authorize this." Without evidence about what the agent was told and what it did, you cannot tell them apart, and neither can the issuer.
What the Protocols Give You as Evidence
The newer payment layers were built partly to answer the "what did the shopper actually approve?" question. Understanding what each one hands you is the starting point for a dispute strategy.
Stripe shared payment tokens: scoped credentials
With ACP checkouts processed through Stripe, you receive a shared payment token (SPT) rather than raw card details. According to Stripe's documentation, each token carries usage limits: a currency, a maximum amount, and an expiration time. The agent sets the maximum amount to match the transaction total. You can retrieve the token's limits along with the card brand and last four digits, and Stripe sends a webhook when a token is deactivated, including when the agent revokes it.
That scope is useful evidence. A charge that stayed within a limit the agent set for that specific order shows the agent asked to pay that amount. Store the token details with the order rather than relying on looking them up months later.
AP2 mandates: signed intent and cart
Google's Agent Payments Protocol (AP2) uses signed mandates. An intent mandate records what the shopper asked the agent to do, and a cart mandate records the specific basket approved before payment. Google frames this chain as an audit trail linking intent, cart, and payment. If you receive mandate data, keep it attached to the order. It is the closest thing available to proof of what the customer approved.
The limitation is worth being honest about: a signed cart proves what was bought, not that it matched what the shopper meant. And if an order did not flow through a mandate-based system, you may not have this evidence at all.
Network-verified agents
Visa's Trusted Agent Protocol and Mastercard's Agent Pay let verified agents identify themselves with cryptographic signatures and pay with agentic tokens. We cover both in our guide to Visa and Mastercard agent payments. For disputes, the point is simple: an order from a verified agent carrying a network token is a much stronger record than an anonymous automated session.
The New Merchant-Side Tools
Vendors have started to treat agent disputes as their own category. Two recent launches show where the market is heading:
- Forter's Agentic Commerce Engine (announced October 2026) combines fraud management for agent traffic, headless checkout orchestration through agentic protocols, and an AI activity dashboard that shows merchants which agents are transacting on their site.
- Chargebacks911's AI agent purchase indemnification (announced late October 2026) identifies agent-made transactions and indemnifies merchants against qualifying chargebacks on them. Eligibility terms and pricing were not detailed in the coverage we reviewed, so read the contract closely. This is a private commercial guarantee, not a change to card network rules.
Neither removes the underlying liability. They either reduce the chance of a bad agent order or shift the cost to a vendor for a fee.
An Evidence Checklist for Agent Orders
You do not need to wait for the networks to publish agent-specific dispute rules. Start capturing this now, because you cannot reconstruct it after a chargeback arrives:
- Tag the channel. Record on every order which agent surface it came from (ChatGPT, AI Mode, Gemini, Perplexity, a verified network agent) so agent orders are reportable and searchable.
- Store the payment scope. Keep SPT usage limits, network token type, and any agent identity data alongside the order record.
- Keep mandate and session data. Retain AP2 mandates, checkout session IDs, timestamps, and the exact product, variant, and price the agent submitted.
- Prove the listing was accurate. Disputes about "not what I asked for" are much easier to win when your product data clearly matched the order. Accurate titles, variants, and specs matter here as much as they do for being chosen in the first place.
- Make policies machine-readable and consistent. Return and shipping terms that agents can read, and that match what you enforce, give you a clean answer to "I wasn't told." See our post on shipping and return policy signals.
- Reduce confusion-driven disputes. Use a recognizable billing descriptor, mention the agent-led purchase in confirmation emails, and make cancellation easy before shipment. Checkout.com recommends similar steps in its guidance on agentic chargebacks.
- Do not blanket-block agents. Blocking agent traffic avoids disputes by avoiding sales. The better move is to recognize verified agents and score the rest, which is the approach the newer fraud tools take.
Where This Is Heading
Expect the rules to shift. Industry commentary increasingly suggests that liability may eventually follow whoever signed off on a cart that deviated from the shopper's intent, which could push some losses back onto agent platforms. That is an opinion about the future, not a rule today.
Until it changes, treat agent orders as your orders. The merchants best placed when the rules do change will be the ones who already have clean agent-order data and accurate listings, and whose checkout accepts verified agents. If you want a review of how your store handles agent traffic end to end, from crawler access to checkout, our technical foundation work starts there.
Want to see how AI engines perceive your brand?
Get an Audit for $87