What is agentic commerce? Discovery, decisions and execution

Agentic commerce describes shopping and commercial workflows in which AI agents interpret a goal and help carry out parts of the journey. That can mean finding suitable products, preparing a purchase or coordinating a merchant task. The useful question is which part the agent handles, whose interests it serves and what it can actually authorize. A conversational recommendation and an executed purchase are different capabilities.

Separate three parts of the buying journey

Discovery helps a buyer find relevant options. Decision support compares those options against needs such as compatibility, price or delivery. Execution changes something: creating a checkout, placing an authorized order or updating a permitted record. A product can support one of these stages without supporting the others. Ask vendors to demonstrate each claimed step rather than treating “agentic” as an all-purpose feature label.

Google's Universal Commerce Protocol walkthrough demonstrates a sample checkout lifecycle. It provides a concrete technical example of an agent interacting with commerce capabilities; a sample is not proof of eligibility or production readiness for a particular merchant. Google's UCP walkthrough. Check current platform, market and integration requirements before planning a deployment.

Identify whose goal the agent serves

A buyer's agent might compare products against a budget and delivery deadline. A merchant's agent might help maintain catalog information or prepare a recovery response. Their goals and permissions are not interchangeable. The merchant's commercial objective does not authorize changing a buyer's preferences, and a buyer's request does not automatically authorize access to the merchant's internal systems.

Consider a fictional replacement-part purchase. The buyer needs a compatible part delivered before Friday. Discovery finds options; decision support checks documented compatibility and delivery conditions; execution proceeds only through an available, authorized purchase path. If compatibility has not been verified, recommending the purchase as though it had been confirmed would conceal a relevant uncertainty. The appropriate next step is clarification or another source, not automatic checkout.

Readiness begins with reliable commercial facts

Review the product identity, purchasable variant, supported attributes, price, availability and fulfillment conditions. Establish which source owns each fact and how changes reach other channels. An integration can transmit inconsistent information as efficiently as correct information. It cannot resolve an undocumented product claim or a promotion whose conditions are unclear.

Use the product-data readiness checklist for a bounded comparison of source, page, feed and checkout. That guide addresses the catalog inspection itself. Start with one family and one market rather than assuming that a successful demonstration represents every product and exception.

Choose a first use case with an observable result

Write the task in a sentence that names the user, inputs, permitted action and evidence of completion. “Help customers shop” is too broad. “Compare documented dimensions for three eligible products and present the source pages” is easier to evaluate. A more consequential task needs stronger authorization and verification, not simply a more persuasive conversation.

What decision is currently difficult for the buyer or operator?
Which facts and systems are needed to support it?
What can the agent propose, and what can it execute?
Which exception returns the task to a person?
What observation will establish that the task finished correctly?

Evaluate the work before discussing revenue

Measure accurate answers, acceptable task completion, review effort and unresolved exceptions. Commercial effects require separate evidence: identifiable visits, qualified actions or orders within a suitable observation design. A working checkout integration does not establish incremental sales, and a recommendation does not demonstrate that a buyer accepted it.

For an ARS conversation with Nextriad, choose one commerce decision and bring the source information and failure cases your team already knows. The next step is to assess a bounded workflow, including what stays with a human. Agentic commerce becomes useful when the task and its limits are clear enough to inspect.