Securafy AI Lab

AI Agent Purchase Liability: Who Pays When It Goes Wrong

Written by Ric Hall | Nov 13, 2026, 4:00:00 PM

If an AI agent buys the wrong thing on the company card, who absorbs the cost: your business, the card issuer, the merchant, or the company that built the agent? Finance leaders ask that question as soon as someone proposes letting an agent reorder supplies, book travel, or renew software. The honest answer is less comfortable than most people expect.

In most cases today, the business that gave the agent its payment credentials carries the loss. U.S. payment law generally treats handing credentials to an agent as authorizing it, card networks have no agent-specific liability shift yet, and agent vendors rarely accept that risk by contract. Your real protection comes from contracts and records, not the chargeback process.

Who is liable when an AI agent makes a bad purchase?

AI agent purchase liability usually lands on whoever authorized the agent to pay, which in a business setting means the company. The principle is older than AI. A June 2026 Goodwin analysis of agentic payment authorization describes disputes between a user and their financial institution as largely settled: a user who gives credentials to a third party, human or AI, is responsible for what follows unless the institution was notified of a revocation or broke its own security procedures.

The same analysis notes that under the Electronic Fund Transfer Act and Regulation E, a transfer is presumed authorized when a consumer furnishes credentials to an agent, even if the agent acts outside what the consumer intended. That is the consumer rule, and it already leans against the account holder. For a business, the starting position is weaker still.

Here is the belief worth challenging. Many owners treat the chargeback as a backstop for agent mistakes, the way they treat it for a stolen card. That instinct was shaped by consumer protections. A business deploying agents is mostly operating outside those protections, and the few that remain were never written for software that is authorized to spend but makes a poor decision.

Why the chargeback is not your safety net

Start with the scope of the rules. The Regulation E definition of an account under 12 CFR 1005.2 covers consumer asset accounts established primarily for personal, family, or household purposes. A company operating account is not one of them. Commercial payment orders and wire transfers fall under UCC Article 4A, where Goodwin notes that a payment order is generally binding on the customer if the bank followed an agreed, commercially reasonable security procedure in good faith, even if the order was unauthorized.

Credit cards look friendlier at first. The Regulation Z liability cap for unauthorized use is the lesser of $50 or the amount obtained before notice. But the definition matters: 12 CFR 1026.12 defines unauthorized use as use by someone without actual, implied, or apparent authority. An agent you configured and funded has, at minimum, a strong claim to apparent authority. The same section also lets an issuer and an organization with 10 or more employee cards agree to liability terms outside the cap, which is exactly how many commercial card programs are written.

The practical consequence is that the clean fraud dispute most leaders imagine rarely fits an agent error. If the agent bought the wrong item, the wrong quantity, or the right item at the wrong price, the card network sees a legitimate, authenticated purchase from an authorized user.

What happens when the agent was authorized, but not for this?

This is the new dispute category, and it does not have a clean home yet. Worldpay's review of agentic commerce liability says a claim that an agent misunderstood an instruction or bought the wrong item is a dispute, not a fraud claim, and that allocation among issuer, acquirer, agent platform, and merchant is still being negotiated. It also points out that Regulation E still assumes a binary distinction between authorized and unauthorized, with no framework for a buyer who authorized the agent but says it misread the instruction.

Worldpay adds a detail that should shape your record-keeping. Without a signed record of what the buyer actually authorized, a disagreement between buyer and agent becomes what it calls a coin flip. For a business, that means the outcome of a dispute depends heavily on whether you can produce the instruction, the limits, and the approval trail, which is the same evidence problem we describe in our piece on why application logs cannot defend an AI agent's decision.

What have card networks and issuers actually changed?

The networks have built identification, not a new liability model. Chargeflow's July 2026 breakdown of AI agent chargeback liability puts it plainly: there is no rule that says the AI provider is liable because an agent made the purchase, and the merchant that sells the goods remains the merchant of record. Authenticated, tokenized agent transactions can move fraud liability to the issuer, but that shift concerns fraud, not complaints about an agent's judgment.

The most concrete step so far came from American Express. In its April 14, 2026 announcement of Amex Agent Purchase Protection, the company said that in the future, if a Card Member authorizes a registered agent and that agent sends Amex the customer's authenticated purchase intent, Amex will protect eligible customers from charges related to agent error, with terms and conditions applying. Read that carefully: it is a forward-looking commitment, it applies only to registered agents passing authenticated intent, and eligibility is not defined in the announcement.

Policy bodies see the same gap. The IMF's April 2026 note on how agentic AI will reshape payments says questions about liability and the legal status of agentic actions remain, and recommends clarifying legal frameworks to reduce uncertainty around liability and enforceability. Until regulators or networks close that gap, the allocation of loss for a business buyer will be decided by contract.

Where AI agent liability really gets decided: your contracts

From a compliance standpoint, this is the part leaders can control. Goodwin's analysis says disputes between a user and the agentic-payment platform are not resolved by the existing framework; common-law principles fill the gap, and outcomes turn on what the user said, what the agent understood, and what records exist. Network compliance may protect a platform from network fines, but it does not stop customer claims for breach of contract or exceeding delegated authority.

That means three documents decide most of your exposure: the agent vendor agreement, your commercial card program agreement, and your insurance policies. In regulated environments, these reviews belong with the same people who already own third-party risk, which is why the discipline in our guide to protecting your business from vendor vulnerabilities applies directly here.

DocumentWhat to look forWhy it matters in a dispute
Agent vendor agreementDefined scope of authority, indemnity for actions outside configured limits, log retention and export rightsYour claim against the vendor depends on proving the agent exceeded the authority you set
Commercial card program agreementLiability terms for employee and non-human cardholders, dispute windows, notice requirementsBusiness card programs can set liability outside the consumer cap
Cyber, crime, and E&O policiesAI exclusions, definitions of authorized user, voluntary-parting and social engineering clausesA loss caused by your own authorized agent may not look like covered fraud
Internal delegation policyWritten approval of which agent may spend, on what, up to what amountIt is your best evidence of what was and was not authorized

The insurance row deserves its own review. We have already seen AI exclusions appearing in business insurance policies, and the pattern resembles what happens with payment fraud generally, where BEC and deepfake losses are often only partly covered because an authorized person moved the money. A purchase made by an agent you authorized can fall into the same gap.

What should your business do before an agent gets a card?

The goal is not to prevent every bad purchase. It is to make sure that when one happens, you can show what was authorized, recover from the right party, and keep the loss small. Operational spend controls do part of that work, and our published guidance on keeping AI in finance to assist and flag rather than move money covers where the line should sit. The liability side adds a short set of requirements:

  • Give each agent its own scoped, tokenized credential rather than an employee's card, so every transaction traces to one agent and one approval.
  • Write down the agent's authority, including spend caps, merchant categories, and approval thresholds, and keep that record with your contract files.
  • Retain the instruction, the agent's proposed cart, the approval, and the payment record for at least as long as your dispute window.
  • Negotiate vendor terms that address errors outside configured limits and guarantee log export if you leave the platform.
  • Ask your broker in writing whether a loss from an authorized agent's mistaken purchase would be covered, and under which policy.

None of this requires waiting for card networks to finish their rules. It does require a deliberate decision about who in the business owns agent purchasing authority, the same way someone owns signing authority on contracts. If that owner does not exist, the default liability allocation described above is the one you get.

How Securafy helps

When a client wants agents to transact, we start with the paperwork that decides liability rather than the agent configuration. That means mapping which payment instruments an agent can reach, reviewing the vendor agreement and card program terms with your finance and legal leads, and checking your cyber and crime policies against the scenario of an authorized agent making an unwanted purchase, using the questions in our guide to what to consider before you buy cyber insurance coverage.

For regulated clients, we then make sure the evidence exists before it is needed: per-agent credentials, retained instructions and approvals, and logs your auditors and your card issuer will accept. The result is an agent program where a bad purchase is a recoverable business event rather than an argument nobody can win.

Where To Go From Here

If your business is about to give an AI agent spending authority, settle the liability questions in your contracts and records first, because the chargeback process will not settle them for you. The time to decide who pays for an agent's mistake is before the first one posts.

If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.

If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.