Businesses Face New Rules For AI Agent Payments

AI agents are placing orders, calling paid APIs and reaching checkout pages. Businesses must decide whether to accept payments from software acting on a customer’s behalf and how to handle purchases made without direct human review.

The payment may still use a card network or a stablecoin transfer that an existing processor can handle. The main changes occur before and after payment. A seller may not know what instructions the agent received, whether a person approved the purchase or who should cover a loss if the agent selects the wrong product.

Merchants can increasingly verify agents with cryptographic signatures instead of relying on user-agent strings or IP addresses. Agents can sign requests using HTTP Message Signatures under RFC 9421. A merchant or security provider can then check whether a request came from a known agent.

Visa and Mastercard maintain directories of agents registered with their networks. Merchants can compare an identifier in a request with those records. An Ed25519 signature can show that the request was not altered. Protocols can also distinguish between browsing and payment attempts and use a one-time value, called a nonce, to help prevent replayed requests.

Fraud liability is clearer when an agent is properly authenticated and a payment token is issued correctly. Existing rules for tokenized card transactions generally apply in those cases. The issuer carries the fraud risk, while the cardholder retains chargeback rights.

The position is less clear when an agent misunderstands an instruction, chooses the wrong product or buys something the customer later says was not approved. Issuers, acquirers, agent platforms and merchants are negotiating how to allocate those losses. No general liability shift exists for these disputes.

U.S. consumer rules generally focus on whether a transaction was authorized. That leaves limited guidance for a transaction in which an agent misunderstood an instruction. Strong customer authentication requirements in Europe could prevent some agent-led payments from being completed.

Card payments keep merchants within existing payment and dispute systems. Customers can still file chargebacks and claim that an agent acted without approval. Visa began piloting its agentic commerce program in April 2026. Mastercard has published an agent acceptance framework designed to work through existing merchant agreements. Availability can differ by acquirer and region.

Stablecoin payments have a different risk profile. Settlement is final and there are no chargebacks, allowing merchants to charge for individual API requests or other small services. That model suits payments between agents and software services, but it gives customers no formal reversal process when an agent makes a mistake or a payment is disputed.

Businesses can set separate rules for browsing and purchasing, including limits on the value and frequency of agent-initiated orders. They can accept verified agents while blocking unidentified automated traffic, record the agent’s request and authorization, and retain evidence of what was delivered.

Paid APIs, data, computing services, digital goods and credits are among the first areas being considered for agent payments because delivery can be checked immediately. Standardized physical goods may be accepted through card payments, which preserve return and dispute processes. High-value, customized, regulated and age-restricted products carry additional exposure because age checks, suitability rules and consumer disclosures often assume that a person is completing the purchase.

Businesses testing agent payments can start with one product or endpoint, set per-transaction and daily spending limits, and require human approval above those limits. They can compare refund and dispute rates for agent orders with rates for purchases made directly by customers before expanding the program.