The Agentic Checkout: Payments for AI Agents

An agent that can research a purchase but cannot complete it saves only part of the work. An agent with unrestricted payment access creates a much larger problem. Useful agent commerce sits between those extremes.

Authority Must Travel With the Payment

Before execution, bind the transaction to:

• The person or organization delegating the purchase.
• The agent identity and task.
• Allowed merchant, category, item, quantity, and delivery destination.
• A maximum amount, currency, and validity window.
• Required evidence and approval conditions.

The payment credential should be narrow, temporary, and unusable for a different purchase. A merchant instruction cannot expand the budget or change the delivery address.

The Agent Needs a Receipt Loop

After payment, verify the final amount, merchant, item, taxes, delivery promise, and return terms. Store the receipt with the task and reconcile it against the original intent. If the result differs, stop the workflow and ask for review.

Subscriptions, variable pricing, tips, deposits, and partial fulfillment need explicit rules. “Under the limit” does not mean “matches the request.”

Preserve Consumer and Business Controls

Users need cancellation, return, dispute, and fraud paths that do not depend on the original agent still running. Organizations need separation of duties, category controls, and audit records compatible with finance systems.

Agentic payments should feel less like handing a bot a credit card and more like issuing a single-purpose purchase order that can execute electronically.