Agentic Commerce: What It Is and What It Changes for Enterprise Buyers

For twenty years, online selling has assumed one thing: a person is looking at your page. Every product grid, every filter, every checkout button was designed for human eyes and a human hand on a mouse. Agentic commerce breaks that assumption. The buyer is still a person, but the thing reading your catalog and pressing the button is software.

That shift is small to describe and large to absorb. This guide covers what the term means, how a transaction actually flows, which protocols and platforms are in play, and what it takes for a business to be ready.

What Agentic Commerce Actually Is

Agentic commerce is a transaction where an AI agent acts on a buyer's behalf: it searches, compares options, makes a selection, and completes the purchase, without the person handling each step.

The word doing the work in that definition is acts. Product recommendation engines have been around for years, and they only ever suggest. A recommendation ends with a human deciding. An agentic transaction ends with the agent deciding and paying. The person sets the goal and the constraints, then steps back.

A useful way to hold the distinction:

• Traditional ecommerce: the customer browses, chooses, and checks out.
• AI-assisted ecommerce: the customer browses with help. A model surfaces options, answers questions, sorts a list. The customer still chooses and checks out.
• Agentic commerce: the customer states an intent, such as reorder our usual filters, but not if the price is above $40 a unit. The agent handles the rest and reports back.

What separates the second from the third is authority, not model quality. Someone has to decide the agent may spend money, and under what limits. That is a policy question before it is a technical one, which is why these projects tend to stall in legal review rather than engineering.

How Agentic Commerce Works, Step by Step

The mechanics are less exotic than the label suggests. A complete transaction moves through five stages.

1. Intent. The buyer gives the agent a goal with boundaries attached: a budget ceiling, a delivery window, an approved vendor list, a spec that must be met.

2. Discovery. The agent needs to find candidate products. This is where most catalogs fail, and we will come back to it. The agent is not reading your rendered page the way a shopper does. It wants structured facts: price, availability, dimensions, lead time, return terms.

3. Evaluation. The agent compares candidates against the constraints. This step is ordinary software logic more than it is intelligence, and it is the step buyers most want to audit later.

4. Authorization. The agent has to prove it is allowed to spend this buyer's money, up to this amount, for this purpose. Payment networks have spent the last year building exactly this, because it is the piece that did not exist before. A card number sitting in an agent's context window is a liability, not authorization.

5. Settlement and receipt. The purchase completes and the agent reports back with what it bought and why. That last clause matters more than it sounds. An agent that spends without producing a reviewable trail is one that no finance team will approve twice.

Stages one through three are where the AI lives. Stages four and five are where the money and the risk live, and they are conventional systems problems.

The Protocols Underneath It

For an agent to buy from a merchant it has never seen, the two sides need a shared language. Several protocol efforts are competing to be that language, and this layer is moving quickly enough that any specific claim here has a short shelf life.

What is stable is the shape of the problem. Any workable protocol has to answer four questions:

• How does a merchant publish what it sells in a form an agent can parse without guessing?
• How does an agent prove it is acting for a real buyer with real authority?
• How does the merchant confirm the money is good before releasing goods?
• Who is liable when the agent buys the wrong thing?

That fourth question is the one with no settled answer. If an agent orders 500 units instead of 50, the chain of responsibility runs across the buyer, the agent's operator, the model provider, and the merchant, and none of them have historically owned that failure. Expect this to be argued in contracts long before it is solved in code.

The practical read for a business: do not bet your architecture on one protocol winning. Bet on your product data being clean and machine-readable, because every candidate protocol needs that, and it is work you have to do regardless of which one prevails.

Agentic Commerce Examples You Can Point At Today

The category is young, so the honest examples are narrow.

Repeat procurement. A distributor's customer reorders the same consumables every month against a standing contract. The spec is fixed, the vendor is approved, the price band is known. An agent handles this well because the decision space is small and the failure mode is cheap. This is where most real deployments are.

Constrained sourcing. A buyer needs a part matching a spec, delivered by a date, under a ceiling. The agent searches approved suppliers and returns a shortlist or completes the order. The value is in the search hours removed, not in cleverness.

Assisted retail baskets. A consumer asks an assistant to assemble an order under a budget. This is the demo that gets the press coverage. It is also the one with the thinnest margin of error, because taste is not a constraint you can express in a spec.

The pattern across all three: agentic commerce works where the decision is bounded and checkable. It struggles where the decision is subjective. If your product is chosen on judgment rather than specification, this wave will reach you late.

Agentic Commerce Platforms: Who Is Building What

Three groups are building agentic commerce platforms, and they want different things.

Payment networks and processors are building the authorization layer. Their interest is straightforward: if agents are going to spend, the spending should run over their rails with their fraud controls. This is the most mature part of the stack because it is closest to what they already do.

Commerce platforms are building the merchant side, exposing catalogs and checkout in agent-readable form. If you sell on a hosted platform, some of this will arrive as a feature you switch on rather than a project you fund.

Model and assistant providers are building the agent itself and want to be the surface where buying intent begins. This is the land grab, and it is why announcements in this space arrive as partnerships between an assistant and a payment network.

For most businesses the useful question is which of these three already sits in your stack, because that is where agentic capability will show up first and cheapest.

What Changes for Your Business

Four things change, in rough order of how soon you will feel them.

Your product data becomes an interface. Today your catalog is decoration around a buying decision a human makes. In agentic commerce it is the decision. Missing dimensions, stale stock counts, and inconsistent units stop being tidiness problems and start being lost orders, because the agent cannot ask a salesperson and will simply not select you.

Your funnel loses its middle. Persuasion assumes an audience. Merchandising, copy, urgency cues, and layout are all aimed at a human who can be moved. An agent filtering on price, availability, and spec is not moved. Whatever share of your conversion came from persuasion is exposed.

Your traffic quality shifts. Fewer sessions, higher intent per session. Reporting built on visits and time on page will describe less and less of what happened.

Your support load moves. Buyers stop asking whether the item is in stock and start asking why the agent chose it. That is a different question and your current support content does not answer it.

None of this calls for rebuilding anything this quarter. It does mean treating structured product data as infrastructure rather than housekeeping, because that work is slow, unglamorous, and required for every version of this future.

A Readiness Check Before You Invest

Before funding an agentic commerce project, answer these honestly.

Can a machine read your catalog without guessing? Not your website. Your data. If price, stock, and spec live in three systems that disagree, no agent will transact with you reliably, and no protocol fixes that.

Is your buying decision bounded or subjective? Bounded decisions are ready now. Subjective ones are not, and no amount of model quality changes that in the near term.

Who signs off on an agent spending money? If you cannot name the person and the limit, you do not have a project yet. You have a demo.

What happens on the first wrong order? Not whether. When. If the answer is not written down, the pilot will end at the first mistake regardless of how well the other 99 orders went.

Is the constraint your technology or your contracts? For most enterprises it is contracts. That is worth knowing before you spend an engineering quarter on the wrong bottleneck.

The category is real and the direction is not in doubt. The timing is very much in doubt, and the businesses that do well here will be the ones that spent the waiting period making their product data honest rather than the ones that shipped an agent first.