Agentic commerce explained: how AI agents buy from B2B sites

TL;DR

• Agentic commerce means AI agents doing part of the buying for a person: research, comparison, forms and, sometimes, checkout.
• In B2B the research half is already here. Gartner found that 45% of B2B buyers used generative AI in a recent purchase, mostly to gather information on vendors and products.
• The transaction half is early. As of October 2026, agent checkout protocols target consumer retail, and B2B features such as quotes, approvals and purchase orders are still open proposals.
• What pays off now is a website that agents can read and use: public facts in the HTML, a clear "how to buy" page and forms an agent can complete.

What agentic commerce means

An AI agent is software that takes steps toward a goal on someone's behalf. In commerce, that goal might be "find three accounting tools that work with our ERP and cost under $50 per user," or "get quotes from two freight forwarders for a monthly lane from Rotterdam to Chicago."

It helps to split the buying journey into two halves:

1. Research. The agent searches, reads vendor websites, compares options and writes a shortlist. It may visit dozens of pages in a few minutes.
2. Transaction. The agent fills in a form, requests a quote, books a meeting or completes a checkout, ideally with the person approving the step that commits money.

Most of what you read about agentic commerce is about the second half: protocols, payment tokens and checkout buttons inside chat apps. For B2B websites, the first half matters far more today.

The research half is already happening

The numbers on AI-assisted research are consistent across the major analyst firms:

• A Gartner survey of 645 B2B buyers, published in May 2026, found buyers used an average of seven information sources in a recent purchase, and 45% said they used generative AI, "primarily to gather information on vendors and products." The same survey found that 69% prefer to validate AI-generated insights with a sales rep.
• Forrester's State of Business Buying 2026 reports that "nearly all business buyers (94%) report using AI during their buying process."
• Cloudflare wrote in September 2026 that daily requests from AI agents on its network grew by more than 1,700% over the past year, and that "an agent booking a table, comparing insurance quotes, or buying a dataset for a researcher is a customer."

The pattern is familiar to anyone who has watched a buyer use ChatGPT or Gemini at work. The assistant searches, reads pricing, integration and security pages, and returns a comparison. Where a page is missing a fact, the assistant either skips the vendor or fills the gap with a guess. Our guide to how ChatGPT, Gemini and Perplexity describe your company covers where those answers come from.

Agents that act in the browser

A newer kind of agent goes further and operates a real browser. These are usually called agentic browsers. Google's auto browse in Chrome, in preview for U.S. Google AI Pro and Ultra subscribers since January 2026, lists early uses such as "filling out tedious online forms" and "getting quotes for plumbers and electricians." Google says auto browse is "designed to pause and explicitly ask for your confirmation" before actions like "making a purchase."

That is the realistic picture of agentic commerce on a B2B site in 2026: an agent that reads your pages, then fills in your quote or demo form while the buyer watches.

The transaction half: protocols as of October 2026

Several companies have published protocols so that agents can transact without guessing at buttons. Here is where each stood when we checked.

ACP and ChatGPT

OpenAI launched Instant Checkout with the Agentic Commerce Protocol in September 2025, starting with U.S. Etsy sellers. In March 2026 it changed course. OpenAI wrote that "the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide," and that it would let merchants use their own checkout while it focused on product discovery. Digital Commerce 360 reported that the checkout experience appeared to be moving into ChatGPT apps, where merchants have more control.

UCP and Google

Google announced the Universal Commerce Protocol on January 11, 2026, for checkout in AI Mode in Search and the Gemini app, starting with eligible U.S. retailers. Google stresses that "Retailers remain the seller of record." The UCP site lists catalog search, cart building, identity linking, checkout and order management, with support for MCP, A2A and AP2.

The B2B gaps are visible in the project's own issue tracker. As of October 2026, these were open proposals:

• Catalog items with no public price, the classic "price on request" product.
• A quote extension for a time-bound price the business issues and the buyer accepts. Its own example is a buyer asking "quote me 500 units."
• A B2B procurement charter covering buyer eligibility, organization policy, business pricing, requisitions, approvals, purchase orders and reconciliation.

AP2, Visa and Mastercard

Google's Agent Payments Protocol, announced in September 2025, uses signed "mandates" to prove a person authorized a purchase. An Intent Mandate captures what the person asked for, and a Cart Mandate is "a secure, unchangeable record of the exact items and price." Visa's Trusted Agent Protocol and Mastercard's Agent Pay solve a different problem: telling a registered agent from an ordinary bot. Cloudflare explains that both build on Web Bot Auth, signed requests that let a merchant "distinguish whether a particular interaction is intended to browse or to pay."

Separately, the WebMCP draft lets a page offer named actions, such as "request a quote," that a browser agent can call directly. Our plain-English guide to WebMCP explains how it relates to the Model Context Protocol.

Why B2B buying is harder to automate

Consumer checkout is one person, one cart and one card. B2B buying rarely looks like that:

• Prices depend on the buyer. Volume tiers, contract rates and negotiated discounts mean there is often no single public price.
• Many people decide. Forrester's 2026 research puts the typical buying decision at 13 internal stakeholders and nine external influencers, and finds procurement teams are decision-makers in 53% of buying cycles.
• Approvals come before payment. A requisition, a manager's approval and a purchase order usually precede any invoice. The UCP procurement charter lists exactly these steps as unsolved.
• Payment is on terms. Net 30 invoices, purchase orders and bank transfers are normal; card checkout is the exception.
• Vendors need onboarding. Security questionnaires, tax forms and contracts often come before the first order.

None of this stops agents from being useful. It means that for most B2B sellers in 2026, the agent's job ends at a well-formed request: a quote request with the right details, a booked meeting or a shortlist with accurate facts.

What agents need from a B2B website

Here is what an agent looks for when it researches you, and what you can do this quarter.

Facts it can read

• Prices or price ranges. If you publish prices, put them in text, not in images. If pricing is custom, say so plainly and give the drivers: "Pricing depends on seat count and data volume; most teams of 20 to 50 users pay between X and Y per month." An agent cannot cite a number you never wrote down.
• Limits and fit. Plan limits, minimum order quantities, lead times, service areas and supported integrations. These are the details that decide whether you make a shortlist.
• Trust details. Certifications, data hosting regions and security documentation, on pages that load without a login.

A clear way to buy

Write a short "how to buy" page that answers the procurement questions agents and buyers ask: self-serve or quote, accepted payment methods, whether you accept purchase orders, invoice terms, contract length and how to request vendor onboarding documents. This is the page an agent will quote when a buyer asks "can we pay by invoice?"

Forms it can complete

An agentic browser fills in forms the way a careful person would, so the basics matter:

• Visible labels on every field, not placeholder text alone.
• Plain field names (company, email, quantity) and standard autocomplete attributes.
• Error messages in text, next to the field.
• A short first step. Ask what you need to route the request; ask the rest later.
• No surprise steps that only a human can pass on public pages. If you need bot protection, apply it to abuse patterns rather than to every first submission.

Access that matches your intent

Decide which AI crawlers and fetchers may read your public pages, and check that firewall rules do not block them by accident. Our checklist for an agent-ready website walks through crawler access, content in HTML and measurement, and the guide to AI agents visiting your website shows how to tell them apart in your logs.

A worked example

Imagine a buyer asks an assistant: "Find two vendors for SOC 2 compliant meeting scheduling for a 40-person sales team. Budget under $20 per user per month. We pay by invoice."

The agent searches, opens five vendor sites and reads each pricing, security and billing page. Vendor A publishes per-seat prices in a table, links to its SOC 2 report request page and states "annual plans can be paid by invoice." Vendor B shows prices in a screenshot and says "contact sales" for billing. Vendor C's pricing page only renders after a script runs.

The agent shortlists Vendor A with confidence, lists Vendor B with "billing terms unclear" and leaves Vendor C out. If the buyer asks it to continue, it opens Vendor A's quote form, fills in team size and billing preference, and asks the buyer to confirm before submitting.

Nothing in that flow needed a new protocol. It needed facts in text, a page about how to buy and a form that works.

Where sources disagree

Analysts agree that buyers use AI heavily for research, but not on how much they trust it: Gartner's survey found that 51% of buyers think generative AI is more likely to give them misleading information than a sales rep. Vendors building protocols, meanwhile, describe a near future of autonomous buying. The evidence so far is more cautious. OpenAI pulled back on in-chat checkout within six months, and the UCP features B2B sellers need most were still open proposals in October 2026. Prepare for agents that research and request, and watch the protocols without rebuilding your checkout for them yet.

If you want a quick read on how your own site looks to an agent, the agent readiness checker runs the basic checks.

Agentic commerce will change how B2B buyers find vendors long before it changes how they pay them. The sellers who benefit first will be the ones whose websites already answer an agent's questions in plain text.