← Articles

Who Is Liable When an AI Agent Buys Something? Authorization and Liability in Agentic Commerce

Published July 29, 2026Updated July 29, 2026Trust

Key takeaways

  • Liability in card payments follows authorization, so the main question is not who clicked but whether the cardholder can be shown to have authorized the purchase.
  • The core gap in agentic commerce is the identity chain problem: a merchant usually sees only the agent that executes the purchase, not the person behind it or the agents in between.
  • A recorded approval of specific line items, confirmed with a passkey or biometric, is strong evidence that a specific person authorized a specific purchase.
  • When the purchase runs on the user's own card, disputes and chargebacks go through the user's bank as normal, and the usual cardholder protections apply.
  • An approval of specific terms gives an issuer more authorization context than a spending cap, because it describes a decision and not only a dollar ceiling.
  • This article explains the general framework. It is not legal advice, and specific outcomes depend on card network rules, issuer policy, regulation, and the agreements between the parties.

When an AI agent buys something, liability follows authorization: if the cardholder authorized the purchase, it is an authorized card transaction, and if they did not, the usual dispute and chargeback process applies. The practical task is to make authorization provable, by having the user approve specific line items with a passkey or biometric and keeping that consent record with the transaction.

This page explains how the authorization framework works in agentic commerce. It is not legal advice. Actual liability allocation depends on card network rules, issuer policy, applicable regulation, and the contracts between the parties involved.

Why is liability confusing when an agent is the one buying?

Liability is confusing for agent purchases because card rules assume that the person who buys and the person who pays are the same person, in the same session. Card payments have a mature liability system, and agentic commerce breaks that assumption without breaking the system.

  • The buyer and the payer are different parties: The agent performs the purchase, and the human owns the card. Existing rules were written for one party in one session.
  • Merchants cannot see the human: A merchant that receives an agent-driven order sees the agent, if it sees anything identifiable. The person who asked for the purchase is not visible to it.
  • The chain can contain several agents: A consumer assistant can delegate to a shopping agent, which calls another service to execute. Each hop loses information about who first asked.
  • There is no standard proof of authorization: A conventional card message does not say that a specific person approved a specific basket. This missing record is the core of what the industry calls the identity chain problem.

The hard part is not deciding who is liable. It is proving who authorized the purchase.

What is the identity chain problem?

The identity chain problem is that, in a multi-party agent transaction, the parties who make a trust decision cannot see the parties whose authority matters. A merchant must accept a purchase without verifiable answers to three questions.

  1. 1

    Who is this agent? Without a verifiable identity, an agent is anonymous automated traffic, which merchants and their infrastructure providers often block by default.

  2. 2

    Who authorized this agent to spend? A recognized agent must still show that a specific person approved this specific purchase, not only that the agent exists.

  3. 3

    What exactly was authorized? A general permission to spend is weak evidence. Authorization is meaningful when it names items, amounts, and a merchant, so it can be checked against what happened.

The three questions belong to different layers. The first is agent identity, which standards such as Web Bot Auth and protocols such as Visa's Trusted Agent Protocol address. The second and third are consent, which a structured approval record addresses.

How can a platform make authorization provable?

A platform makes authorization provable by turning the user's request into structured purchase terms, getting the user's explicit approval of those terms, and recording that approval before anything is bought. The industry calls this record an intent object or a mandate. It works as a receipt issued before the purchase.

  • Specific terms: Items, quantities, prices, fees, the total, the merchant, and fulfillment details such as an address or reservation time. It is a described purchase, not a spending allowance.
  • Native approval: The approval appears inside the product's own app and is confirmed with a passkey or biometric, in the pattern of an Apple Pay confirmation.
  • A logged record: The system stores the approval with the terms, which turns a UI interaction into evidence that parties can refer to later.
  • Enforcement during execution: The system checks each execution step against the approved terms, so a difference between what was approved and what was bought does not pass silently.
  • No exposed card number: The agent works with a tokenized credential, not with a card number it could reuse.

The Agent Payments Protocol (AP2) is one public example of this approach. Google created AP2 and donated it to the FIDO Alliance in 2026. AP2 defines signed mandates: a Checkout Mandate authorizes the completion of a checkout, and a Payment Mandate authorizes the payment for that checkout.

How does that consent reach the bank?

Consent reaches the bank when authorization information travels with the transaction through card network programs for agent-initiated payments. A consent record that stays only in a platform's database helps less in a dispute than one that the network and the issuer can see.

  1. 1

    The purchase runs on the user's own card: The transaction posts to the account the user already holds with their own issuer, so it appears on the same statement and in the same bank app.

  2. 2

    The network recognizes the transaction as agent-initiated: Programs such as Visa Intelligent Commerce and Mastercard Agent Pay use agent-specific tokens, so an agent purchase does not look like unattributed card-not-present activity.

  3. 3

    The executing agent is identifiable: Web Bot Auth and Visa's Trusted Agent Protocol let a merchant verify which agent sent a request, so the executing party is named and not anonymous.

  4. 4

    Disputes follow the normal path: Because the transaction is an ordinary card transaction on the user's own card, the user disputes it through their bank, and the usual cardholder protections apply.

Why is an approved intent better evidence than a spending cap?

An approved intent is better evidence than a spending cap because it records a decision about a specific purchase, while a cap records only a ceiling. Limits are useful controls, but as authorization evidence they are thin.

Question an issuer or merchant asksA spending cap answersAn approved intent answers
Did a person approve this?No signalYes, with a passkey or biometric confirmation
Approve what, exactly?An amount ceilingSpecific line items, total, and merchant
Is this merchant expected?Usually not recordedThe merchant is named in the approved terms
Does the charge match the approval?Only that it is under the capThe charge can be checked against the approved terms
Can the agent exceed the approval?Up to the cap, on anythingNo, execution is limited to the approved terms

The difference matters most in the common case: a legitimate purchase that a risk model finds unusual. A described purchase that the user confirmed is a stronger basis for an authorization decision than an anonymous charge under a limit.

Who is responsible in the common failure cases?

In the common failure cases, responsibility depends on what was authorized, what was executed, and whether the two match. This is a framework, not legal advice, and most cases fall into one of five patterns.

  • The user approved it and got what was described: This is an ordinary authorized card purchase. A complaint about the product or the merchant is a merchant dispute, handled through the usual card channels.
  • The user approved one thing and got another: When the approved terms are stored and checked during execution, the mismatch is visible in the record and does not depend on competing memories.
  • The user never approved anything: With no consent record, the transaction is unauthorized in the conventional sense. The user disputes it through their bank, and cardholder protections apply.
  • The agent misunderstood the request: A separate approval step that shows line items catches the error before the purchase, not on the statement.
  • Someone took over the user's account: Account takeover is an authentication problem. An approval bound to a passkey or biometric on the user's device resists a hijacked session.

What does this mean for an agent builder's risk exposure?

For an agent builder, risk exposure falls when every purchase has a recorded user approval, runs on a tokenized card, and stays on the user's own card and bank. A builder can build the consent, credential, and network pieces in house, or use infrastructure that provides them.

  • Keep card data out of your systems: Use tokenized credentials from a PCI DSS compliant provider, so neither the agent nor your servers see card numbers.
  • Keep a defensible consent record: Store the approved terms and the confirmation for every purchase, so you can answer when a user, a merchant, or a partner asks what happened.
  • Use the networks' agent programs: Agent-specific tokens and verifiable agent identity give issuers and merchants the context to approve legitimate purchases.
  • Do not ask users to trust something new: A charge on the user's own card, with their own bank, avoids a new stored-value account and an unfamiliar payment brand in the middle of the purchase.

Where is this heading?

The industry is moving toward an explicit identity chain: verifiable agent identity at the request layer, agent-specific credentials at the network layer, and machine-readable consent that links a human decision to a specific transaction. All three must be present at once, because any one alone still leaves a merchant without a full answer.

Until that is universal, the practical approach is to run agent purchases through systems that carry authorization to the networks. Then liability falls where it has always fallen for card payments, and not in an unmapped space between an agent, a platform, and a merchant.

Most liability questions in agentic commerce are authorization questions. When authorization is specific, confirmed by the user, recorded, and visible to the networks, an agent-initiated purchase behaves like an ordinary card purchase.

Frequently asked questions

Who is liable if an AI agent makes a purchase I did not want?

Liability follows authorization. If you did not approve the purchase, it is an unauthorized card transaction, and you dispute it through your bank under your normal cardholder protections. Exact outcomes depend on network rules, issuer policy, and regulation, so this is general information and not legal advice.

What is the identity chain problem in agentic commerce?

The identity chain problem is that a merchant usually sees only the agent that executes a purchase, not the person who requested it or the intermediate agents. A conventional card transaction cannot prove that a specific person approved a specific basket.

Do normal chargeback rights apply to purchases made by AI agents?

Yes, when the purchase runs on the user's own card through the card networks. The user disputes it through their bank as with any other card purchase, and the usual cardholder protections apply.

How can a platform prove that a user authorized an agent purchase?

A platform can prove authorization by turning the request into structured purchase terms, with items, total, and merchant, and having the user approve them with a passkey or biometric. The platform stores that approval as a record and checks execution against it.

Why is an approved intent stronger than a card spending limit?

A spending limit shows only that a charge fell under a ceiling. An approved intent shows which person approved which items, at what total, from which merchant, and the charge can be checked against it. That gives issuers more context for an authorization decision.

Is the agent platform liable for what an agent buys?

It depends on the case. Allocation of liability between a user, an agent platform, a merchant, and an issuer depends on card network rules, issuer policy, regulation, and the agreements between the parties. A recorded user approval of specific terms makes it much easier to establish what was authorized. This is general information, not legal advice.

Related articles