Agentic checkout is the completion of a purchase by an AI agent acting on a shopper's behalf, inside the assistant's own interface, rather than by the shopper navigating to the merchant's storefront. The customer expresses an intent — a product, a constraint, a budget — and the agent selects, authorizes and places the order against the merchant's systems.

It is the transactional end of agentic commerce. Where agentic commerce describes the broader shift toward AI intermediating discovery and purchase, agentic checkout is the specific moment money moves.

How the transaction actually works

Three things have to be true for an agent to complete a purchase, and each is a distinct piece of infrastructure:

The agent can read the catalog. Product availability, variants, price and shipping have to be machine-readable and current. A feed built for human browsing is not sufficient.
The agent can construct and submit an order. This requires a programmatic path to cart and checkout that the merchant has explicitly exposed — not an agent driving a browser session and clicking buttons, which is brittle and which most merchants actively block.
Payment can be authorized on the shopper's behalf. This is the hard part, and the reason the protocols exist. The merchant needs assurance that the agent is genuinely authorized by the cardholder, and the shopper needs a bounded mandate rather than an open-ended one.

The protocols involved

Several competing and overlapping standards address the authorization problem. The Agentic Commerce Protocol (ACP) defines how an assistant passes a completed order and payment credential to a merchant. Google's Agent Payments Protocol (AP2) addresses the mandate question — proving that a human delegated a specific, bounded purchase to a specific agent. The Model Context Protocol (MCP) is the more general plumbing by which agents call external tools, including commerce ones.

None of these has won. A merchant deciding today is choosing which surfaces to support rather than adopting a settled standard, and that is a reasonable argument for supporting the surfaces where the customers actually are rather than all of them.

What changes for the merchant

Agentic checkout removes the storefront from the purchase, which breaks several assumptions at once:

Merchandising loses its surface. Upsells, cross-sells, bundles and urgency mechanics live on pages the agent never renders. Order value has to be influenced through catalog structure and pricing rather than through layout.
Product data becomes the storefront. If the agent's only view of a product is its structured data, then attributes, variant naming, and availability accuracy are doing the work that photography and copy used to do.
Attribution gets harder. The order arrives from an assistant, not from a session with a referrer and a campaign parameter. Standard analytics will under-credit the channel unless the merchant deliberately instruments it.
Returns and service assumptions shift. A customer who never visited the store has no relationship with its policies, its support channels or its brand — which raises both return rates and the value of a strong post-purchase experience.

Whether it matters yet

Honestly: for most brands, not much in absolute volume today. Agentic checkout is early, the protocols are unsettled, and the share of orders arriving this way is small for all but a handful of categories.

What makes it worth attention anyway is that the preparation is not agentic-specific. Clean structured product data, accurate real-time inventory, and a machine-readable catalog are the same investments that improve marketplace feeds, shopping ads and AI visibility. A brand that does that work is better off regardless of whether agentic checkout arrives quickly — which makes it one of the few AI-adjacent bets with a defensible downside.

Getting a store ready for that — catalog structure, checkout exposure, and the protocol surfaces worth supporting — is what agentic commerce setup covers, and it sits inside the broader AI-ready ecommerce work.