Agentic Commerce Explained: AI Agents That Buy Things

The customer who never sees your homepage

Picture a customer who asks an assistant to find a waterproof jacket under a set price, in stock in their size, delivered by Friday, from a shop with free returns. The assistant reads a dozen product catalogues, rules out most of them for vague delivery information, and presents two options. The customer taps approve. Your beautiful photography and carefully written homepage played no part at all.

That is agentic commerce in its consumer form, and in 2026 it has moved from demo to early reality through payment networks, large marketplaces and AI assistant providers all adding purchase capabilities. The B2B form is quieter and arguably more important: procurement agents reordering consumables, comparing supplier quotes and raising purchase orders.

How an AI agent actually buys something

1. Intent and limits. The person states what they want and sets boundaries: budget, preferred brands, delivery date, whether the agent may pay or only propose.
2. Discovery. The agent searches, reads product feeds, structured data and pages, and sometimes queries a merchant's API or agent directly.
3. Comparison. It filters on the hard constraints first, then ranks on price, reviews, returns policy and delivery.
4. Authorisation. For payment, networks increasingly issue agent-specific tokens or credentials with spending limits, so the agent never holds the real card details.
5. Checkout. The agent completes the purchase through an API, an agent-friendly checkout, or a browser session, and records what it did.

Step four is where most of the serious engineering is going. Card schemes and payment providers are building ways to prove a purchase was authorised by a real person for a specific agent within a specific limit, which matters enormously for fraud rules and chargebacks.

What changes for online sellers

An agent does not browse. It reads. That shifts value from persuasion to precision.

If a product feed says a jacket is waterproof but not its size range or fabric, an agent filtering on those facts drops it. Nobody sees the product to be persuaded. This overlaps closely with generative engine optimisation, and the fixes are the same: structured data, clean feeds and pages that state facts plainly. Our note on ecommerce structured data is a sensible place to start.

B2B agentic commerce is where the volume is

Consumer shopping agents attract the headlines, but a distributor's customers reordering through procurement agents is the more likely near-term change for SME sellers. Repeat orders for known items, within agreed price lists, with delivery windows that matter more than brand, suit agents well.

Take a wholesaler of catering supplies with 800 trade accounts. If a tenth of them start ordering through agents, the wholesaler needs account-specific pricing an agent can query, stock that is accurate by the hour, and an ordering API that respects credit limits. Most of that sits in the ERP already. The work is exposing it safely.

The risks for buyers and sellers

• Wrong purchases. Ambiguous instructions produce confident wrong orders. Sellers will see more returns that start with the words "my assistant ordered".
• Fraud and disputes. Who authorised the purchase becomes a harder question, and chargeback rules are still adapting.
• Manipulation. Product pages containing hidden text aimed at steering agents are already appearing. Reputable platforms will penalise it, and buyers' agents are being hardened against it.
• Margin pressure. Agents compare ruthlessly on stated criteria. Differentiation has to show up in data an agent can read, such as delivery speed, warranty or bundled service.
• Consumer law. Distance selling rights, cancellation periods and price display rules still apply when an agent buys. The seller's obligations do not shrink.

What to do this year, in order

1. Audit your product data for completeness: every attribute a buyer would filter on, in structured form
2. Make stock and delivery information accurate and available in feeds, not only on the page
3. Publish returns, warranty and shipping terms clearly and consistently
4. Keep checkout short and free of surprises, and make sure it works without scripts that confuse automated sessions
5. For B2B, expose account pricing, stock and order status through an authenticated API
6. Watch your payment provider's agent payment features and adopt them when your customers actually use them

Items one to four help ordinary shoppers and search engines as well, so there is no wasted effort if agentic commerce grows more slowly than predicted.

How we approach agentic commerce readiness

At SpiderHunts we treat agentic commerce readiness as a data and integration job rather than an AI project. The first thing we check is how an agent actually sees the catalogue, by running a shopping-style agent against the site and recording what it can and cannot determine. The gaps are almost always boring: missing attributes, stale stock, inconsistent delivery promises.

For B2B sellers the next step is usually an ordering API with proper account permissions, built as part of our custom software work, with an agent-friendly layer added only when a customer wants one. If you are on the buying side, the same discipline applies in reverse, which we cover in AI agents for procurement approvals.

Frequently asked questions