Who Pays When an AI Agent Overspends?

Companies that deploy AI agents usually bear the cost when the systems spend more than authorized. An AI agent cannot be a legal party, so a dispute generally concerns whether the deploying company or another party is responsible for the transaction.

The issue can take several forms. An agent may repeatedly buy services it was allowed to use, place an order outside its instructions, or have its credentials compromised and used for someone else’s benefit. These situations raise different legal questions. Ordinary overspending is generally treated differently from fraud.

If an agent repeatedly buys permitted services, the deploying company will generally be responsible. For example, an agent that makes tens of thousands of paid API calls may have exceeded its expected usage, but each transaction was within the type of activity the company authorized.

A purchase outside the agent’s instructions depends in part on whether the vendor could reasonably believe that the agent had authority to act for the company. A company email address, access to a procurement system, purchase-order numbers and a history of similar transactions can make an order appear legitimate.

The absence of human review may not invalidate the transaction. In the United States, electronic transaction laws allow contracts to be formed through automated systems without a person approving each transaction at the time it occurs.

An internal spending cap may limit an agent’s activity without determining whether a transaction is binding. If a company limits an agent to $25,000, but a vendor that does not know about the limit receives a routine-looking $30,000 order, the company may still be responsible. A cap is more likely to affect the dispute when it appears in a vendor contract or was otherwise communicated to the vendor.

A company’s conduct after discovering an error can affect the dispute. Keeping, deploying or using goods that arrived may amount to accepting the purchase. Delays in rejecting the order can strengthen the vendor’s position. Legal results vary by jurisdiction, and courts are still developing rules for autonomous agents.

The payment method affects the available recovery options. Card transactions generally retain existing protections for unauthorized or fraudulent payments, subject to the applicable card and token rules. An agent that spends too much or buys the wrong item may fall into a less settled area involving the business, payment providers, platforms and merchants.

Stablecoin payments carry a different risk. Completed transactions are generally irreversible and do not have a standard chargeback process. If an agent sends stablecoins to the wrong recipient or spends more than intended, recovery usually depends on the recipient returning the funds or on other legal remedies.

A limit in an agent’s prompt or configuration may be difficult to enforce against a vendor. A limit applied directly to the card, wallet or account holding the funds can stop a transaction before completion and create an external record of the refusal.

Companies can reduce exposure by assigning a separate card or wallet to each agent or project, setting transaction and daily limits, restricting merchants and requiring human approval for large orders, new vendors, indemnities and automatic renewals. They should retain records of the agent’s instructions, permissions, approvals and transactions.

After discovering an unauthorized purchase, a company should stop further transactions, notify the vendor in writing, avoid using the goods or services and preserve relevant records. Contracts with AI platform and model providers may limit recovery because they often exclude consequential damages or assign responsibility for automated actions to the customer. Insurance coverage may differ depending on whether the transaction involved fraud, authorization or an automated error.