Your AI Agent Bought the Wrong Thing. Who Pays?
If an AI agent buys the wrong thing, a credit card usually lets you dispute it and a debit card often does not. Who pays in five scenarios, under US rules as of November 2026.

When an AI agent buys the wrong thing, who pays depends on three things: how you paid, what you actually authorized, and whether anyone can prove it. In the US, a credit card gives you a legal right to dispute goods "not delivered as agreed," which usually puts the loss on the merchant. A debit card or bank transfer gives you much less, because handing an agent your credentials generally counts as authorizing it. Outright fraud is still covered by the usual rules.
This article explains how existing rules apply as of November 2026. It is not legal advice, and I am not a lawyer. For a specific dispute, talk to your card issuer or a consumer protection attorney.
Key facts
- Agents are buying for real. Visa went live with agent-initiated payments with over 30 European issuers in July 2026. Amazon's Alexa for Shopping can auto-buy at a target price and buy from other merchants' sites for you.
- No new liability rules exist yet. US consumer payment law (Regulation E and Regulation Z) was written with a human pressing "buy." I found no CFPB guidance on AI agents as of early November 2026, and Worldpay's analysis says plainly: "No liability shift exists yet."
- Handing an agent your credentials usually counts as authorizing it. Under Regulation E, a transfer by someone you gave access to is not "unauthorized," even if the agent goes beyond what you meant.
- Your payment method matters more than ever. Credit cards carry the strongest dispute rights. Amex has gone further, pledging to protect cardmembers from erroneous purchases by agents registered with Amex.
- Merchants absorb most losses today. 93% of merchants say AI or agent providers should bear the loss when an agent makes an incorrect purchase, according to PYMNTS Intelligence data cited by Forkast.
Why this question is no longer hypothetical
Picture two cases. You ask your agent to reorder printer ink, and it buys the wrong cartridge, from a seller with a no-returns policy, for $186. Or you ask for a carry-on bag under $150, and it buys a $149 bag plus a $79 "matching" toiletry kit, because a line on the product page claimed the kit was required for checkout. Both are illustrations, not real cases.
A year ago these were thought experiments. Now Visa, Mastercard, and American Express all run systems that let AI agents pay. Amazon, Google, and Meta all ship agents that can complete purchases, as I laid out in my agentic commerce overview. The holiday season is the first big test, and I wrote a seven-rule holiday shopping guide for the practical side.
The rules for when a purchase goes wrong have not caught up. Almost every protection consumers rely on assumes a person clicked "buy."
I founded LoginRadius in 2013 and scaled it to serve more than a billion users. Strip away the payments vocabulary and this is an identity problem: proving who acted, on whose authority, and within what limits. Here is how it plays out today.
Five scenarios, and who usually pays today
| Scenario | What happened | Who usually pays today |
|---|---|---|
| 1. Fraud | Someone stole your credentials or hijacked your agent | The card issuer, under existing fraud rules |
| 2. Agent misunderstood you | Wrong size, wrong model, wrong item | Often the merchant, if you paid by credit card and dispute it |
| 3. Agent exceeded your limits | You said $300, it spent $900 | Unresolved; depends on proof of your instructions |
| 4. Agent was manipulated | A product page or seller tricked the agent | Unresolved; likely the merchant or the agent platform |
| 5. You changed your mind | "My AI bought it, not me" | You, if the merchant can show you authorized it |
Scenario 1: Fraud
If a criminal steals your card or takes over your agent account and makes purchases, you are generally protected the same way you would be for any other card fraud.
When an agent authenticates properly and the payment token was issued correctly, "liability generally follows existing authenticated tokenized-transaction rules," according to Worldpay's analysis. The issuer carries the fraud risk.
The US limits, as summarized by the FTC, are well established:
- Credit cards: the most you might owe for unauthorized charges is $50.
- Debit cards: $50 if you report within two business days of learning of the loss, up to $500 after that, and potentially everything taken if you report more than 60 days after your statement is sent.
Bottom line: fraud is the clearest scenario. Report it fast.
Scenario 2: The agent misunderstood you
This is the most common case and the one the rules handle worst. The purchase was authorized. You gave the agent access, and it bought something. It just bought the wrong thing.
Regulation E, which covers debit cards and bank transfers, assumes a transfer was either authorized or not. Its definition of an unauthorized transfer excludes one made by a person you furnished the card or credential to, unless you told your bank that access was revoked. A June analysis from the law firm Goodwin puts it plainly: "a transfer is presumed authorized if a consumer furnishes credentials to authorize an agent to make payments."
Credit cards give you more room. Regulation Z section 1026.13 treats a charge as a billing error when the goods were "not accepted by the consumer" or not delivered as agreed. You have 60 days from the first statement showing it to send a dispute. As Goodwin notes, when the issuer credits your account, "in practice, it will be the merchant that bears the loss, not the financial institution."
On the card network side, these disputes use existing reason codes. Chargeflow's mapping puts a wrong item under Visa 13.3 or Mastercard 4853, the "not as described" family, and notes the published code lists have no agent-specific reason code.
Bottom line: if your agent paid with a credit card, you likely have a path to dispute a wrong purchase. If it paid with debit or a bank transfer, your position is much weaker.
Scenario 3: The agent exceeded your limits
You told the agent to spend no more than $300. It spent $900.
This is where liability is still being written. Goodwin points out that under Regulation Z, a card payment is authorized if the person who initiated it had "actual, implied, or apparent authority to act" when making it. Goodwin calls that standard broad enough to arguably cover an agent acting outside its intended scope. Under Regulation E, handing over credentials creates the presumption above.
What decides these cases in practice is evidence. Can you show what you actually told the agent? Can the merchant show what the agent was authorized to do?
That is exactly the problem the new payment protocols are designed to solve:
- Google's Agent Payments Protocol (AP2) uses signed "mandates." In Google's description, an Intent Mandate records your rules of engagement, such as price limits and timing. A Cart Mandate records the exact items and price you approved. Google calls the chain "a non-repudiable audit trail."
- Mastercard's Verifiable Intent, co-developed with Google, is described by the FIDO Alliance as "a cryptographic credential framework that transforms user authorization into portable, verifiable evidence." Both it and AP2 were contributed to FIDO in May 2026.
- Visa requires that a provider of agentic payments "must be registered with the network and establish a defensible chain of authorization," according to Goodwin's analysis.
If a signed mandate says $300 and the agent spent $900, you have a strong case. If there is no record, you are arguing your word against a log you cannot see.
Your dispute with the agent platform itself is a separate matter. Goodwin says that is "where the real risk lives": claims against a platform for breach of contract or exceeding delegated authority are ordinary contract disputes. Read the platform's terms before you connect a payment method.
Scenario 4: The agent was manipulated
An agent that reads web pages can be steered by text on those pages. Hidden instructions telling an agent to add an item, pick a premium option, or ignore a budget are a form of prompt injection, the top risk on OWASP's list for AI applications.
No payment rule addresses this directly. If a seller deliberately planted the instruction, that looks like fraud by the merchant, and a credit card dispute is the natural remedy. If a third party manipulated a page, the answer is unclear.
My read (opinion): this is the scenario most likely to push agent platforms into sharing liability. An agent that can be talked into spending your money by a product page is a defective product, and I expect courts and regulators to eventually treat it that way.
Scenario 5: You changed your mind
Merchants are already bracing for a new kind of friendly fraud: "My AI bought that, not me."
Worldpay lists a false "my agent went rogue" claim as friendly fraud. It warns: "Without evidence tying the agent's action to a specific authorized instruction, merchants have no way to challenge it." Where that evidence exists, such as a signed mandate or a confirmed approval, the merchant can win.
Bottom line: if you approved the purchase, an AI agent does not create a new right to return it. Normal return policies apply.
Your payment method is now your protection plan
Based on the current rules, the payment method you connect to an agent matters more than almost any other choice. The rankings below are my reading of US rules, not a legal guarantee.
| Payment method | Protection when an agent gets it wrong |
|---|---|
| Amex card, registered agent | Strongest. Amex pledges to protect cardmembers from erroneous purchases by agents registered with Amex |
| Other credit cards | Strong. Regulation Z disputes for goods not accepted or not delivered as agreed |
| Debit card | Weaker. Regulation E treats transfers by someone you gave access to as authorized |
| Bank transfer or wallet balance | Weak. Few dispute rights for authorized payments |
| Stablecoins and crypto | Minimal. Generally no chargeback mechanism |
American Express made the boldest move. Fortune reported that Amex pledged "to protect users from erroneous transactions made by agents that are registered with Amex" in April. Its reasoning, per Fortune: it already handles disputed transactions at scale, agents will improve over time, and trusted agent payments will bring more volume. The article gave no definition of "erroneous" and no caps or exclusions, so read the terms before relying on it.
Outside the US
- United Kingdom: under regulation 67 of the Payment Services Regulations 2017, a payment counts as authorized when the payer consented in the form and procedure agreed with its provider. Under regulation 77, the payer's share of unauthorized losses is capped at £35, unless the payer acted fraudulently or with intent or gross negligence. HM Treasury's consultation on modernising payment services regulation, which names agentic payments among the changes the rules must account for, closed on October 6, 2026. The government has not yet published its response.
- European Union: most online card payments need Strong Customer Authentication (SCA), meaning proof of the customer with two factors. When Visa took agent payments live in Europe, it used Visa Payment Passkeys to tie each agent-initiated payment to a verified user instruction, which helps issuers meet SCA.
What the industry is building
The pieces of a liability framework are appearing, but not yet as one system:
- Card networks are focused on identifying agents and proving intent: Visa's Trusted Agent Protocol and Payment Passkeys, Mastercard's agentic tokens and Verifiable Intent.
- EMVCo, the standards body behind chip cards, has circulated a draft agentic payments framework. Industry feedback closed September 30, 2026, according to Chargeflow's regulation tracker. Forkast reports that the draft leaves liability for non-fraud disputes unresolved.
- Six banks (Bank of America, Capital One, Commonwealth Bank of Australia, ASB, ING and NatWest) published shared principles for trusted agentic commerce in September. They cover transparency, safety, privacy and data, choice, and interoperability. Bank of America's Mark Monaco named identity, authorization, fraud prevention, liability management and customer protection as areas that need careful approaches.
- Regulators are watching but have not acted. In a September 29 speech at Sibos, Federal Reserve Governor Christopher Waller said "the biggest barrier to scaling agentic commerce" is building sufficient trust among buyers and sellers. He listed authentication, liability and fraud as open questions.
As Worldpay summarized: "No liability shift exists yet."
How to protect yourself today
- Connect a credit card, not a debit card or bank account. It gives you the strongest dispute rights.
- Set a hard spending limit per purchase, and prefer agents that enforce it through the payment credential, not just a setting in the app.
- Keep approval on. Approving each purchase is your best evidence that something unapproved should not have happened.
- Use registered agents on recognized networks. Amex's protection, for example, applies only to agents registered with Amex.
- Save the record. Screenshot or export your instructions and the confirmation for anything significant.
- Check the return policy before approving. "Final sale" is a much bigger problem when you did not pick the item yourself.
- Review your statements weekly while agents have access, and dispute errors quickly. The 60-day credit card window runs from the statement.
- Know how to revoke access. If an agent misbehaves, revoke its payment token or card immediately and tell your issuer. Under Regulation E, telling your bank that access is withdrawn is what makes later transfers unauthorized.
What merchants should do
- Accept verified agents and capture the evidence. Signed mandates, agent identity, and the user's approval are your defense in disputes.
- Make return and cancellation terms machine-readable so agents see them before buying.
- Distinguish good bots from bad ones at the edge. Verified agent protocols exist so you do not have to block all automation.
- Watch for manipulation on your own pages. Hidden instructions aimed at agents, even from third-party content, could become your liability.
For the identity side of this, the Agent Security Map tracks the vendors building agent identity and authorization controls, and I wrote about why most agents have no idea who authorized them.
My predictions (opinion)
- Liability will follow the evidence of intent. Whoever holds the signed record of what the user authorized will win disputes. That makes mandates and verified intent the most important part of agentic commerce, more than the AI itself.
- Agent platforms will be pulled into liability. Today merchants absorb most agent mistakes. I do not think that lasts, and I expect network rules or contracts to start assigning losses to the platform whose agent made the error.
- Credit cards will dominate early agentic commerce, because they carry the dispute rights consumers will want before trusting an agent with their money.
- Card networks will add an "agent error" dispute category. Today there is no clean reason code to separate an agent's mistake from a cardholder-authorized purchase. I expect that gap to close.
- Regulators will clarify that using an agent does not waive your rights. The Center for Data Innovation argued in March that the "CFPB should clarify that consumer-authorized agents do not waive all error resolution rights," in a post on agentic commerce regulation. My guess is that question gets answered within two years.
I go further on where agents and identity are heading in my identity and AI security predictions for 2027.
The bottom line
Today's payment protections were built on a simple assumption: a person decided to buy. AI agents break that assumption, and the rules are catching up one piece at a time.
Until they do, three things decide who pays when your agent gets it wrong: how you paid, what you authorized, and whether anyone can prove it. Use a credit card, set real limits, keep approvals on, and keep records.
FAQ
Can I dispute a purchase my AI agent made by mistake?
Often yes, if you paid by credit card. In the US, Regulation Z lets you dispute charges for goods not accepted or not delivered as agreed, within 60 days of the statement, and in practice the merchant usually bears the loss. Debit cards and bank transfers are weaker, because Regulation E treats transfers by someone you gave credentials to as authorized.
Who is liable when an AI agent makes an unauthorized purchase?
For outright fraud, such as stolen credentials, existing card fraud rules apply and the issuer generally carries the loss. For agent errors short of fraud, liability is still unresolved. As of November 2026 I found no CFPB guidance on AI agents, no network liability shift for agent purchases, and EMVCo's draft framework reportedly leaves non-fraud disputes open.
Does American Express protect purchases made by AI agents?
In April 2026, American Express pledged to protect cardmembers from erroneous transactions made by AI agents registered with Amex, as reported by Fortune. The report gave no caps, exclusions or definition of "erroneous," so cardmembers should review the terms.
Is there a chargeback reason code for AI agent mistakes?
Not yet. Agent disputes use existing codes, such as Visa 13.3 or Mastercard 4853 for an item that is not as described, or fraud codes when the agent was hijacked. Published Visa and Mastercard code lists show no agent-specific reason code.
How can I protect myself when using an AI shopping agent?
Connect a credit card rather than a debit card or bank account, set a hard per-purchase limit, and keep approval for each purchase turned on. Use agents registered with your card network, save your instructions and confirmations, check return policies before approving, and know how to revoke the agent's payment access.
More like this
All AI Security & Agents- Scams & CybersecurityBefore You Let AI Do Your Holiday Shopping: 7 Rules From a Security ExpertBefore an AI agent shops for you, use a credit card, cap its spending, keep approval on and name the stores it may use. Seven rules for…
- AI Security & AgentsEveryone Bought AI Observability. Nobody Owns Agent Behavior.The AI observability market split into four segments (infra telemetry, dev tooling, runtime security, autonomous remediation) with a…
- GEO & AI SearchYou're Optimizing for the Wrong AI Engine. And It's Costing You Enterprise Deals.Two cybersecurity companies told me they're optimizing for Perplexity. Their buyer? Enterprise CISOs.
Get new AI Security & Agents writing
Enjoyed this? Subscribe and tell us what you read most. AI Security & Agents is already ticked for you. No tracking pixels, unsubscribe with one click.