What agentic commerce means, the protocols behind AI agents that buy (ACP, UCP, AP2 and MCP), where the platforms stand, what product data agents need, and why B2B is harder.
Read the full write-up: What is agentic commerce?
Chapters
- 0:00Intro
- 0:26The short definition
- 0:50Why it matters now
- 1:21What an agent does when it shops
- 2:13ACP and UCP checkout protocols
- 3:16AP2, MCP and card networks
- 4:08Which protocol to bet on
- 4:21Where the commerce platforms stand
- 5:18Why B2B is harder
- 5:36What the protocols cover for B2B
- 6:11Buyers won't wait
- 6:30Fix your product data first
- 6:59What a B2B seller should do now
- 8:04Wrap-up
Shorts from this video
Forrester: 1 in 5 B2B sellers would negotiate with buyers' AI agents in 2026 (0:19)
What ChatGPT's product feed requires: start with these 8 fields (0:29)
Why AI agent checkout is not ready for B2B buying yet (0:18)
ACP, UCP, AP2 or MCP: which agentic protocol should you pick? (0:13)
What is agentic commerce? The short definition (0:24)
Transcript
Intro
Agentic commerce is buying and selling where AI agents do the searching, the comparing and, more and more, the paying. They act on behalf of a person or a business, often without anyone clicking through each step. In this video, we'll cover what an agent actually does when it shops, the protocols behind it, where the commerce platforms stand, and why B2B is harder. Then we'll finish with what a B2B seller should fix first.
The short definition
Here's the short definition. AI agents that find, compare, negotiate and buy for someone. To do that, they rely on three things. The first is product data a machine can read. The second is a checkout API the agent can call. And the third is payment credentials the buyer approved. IBM and Salesforce describe it the same way.
Why it matters now
Agentic commerce stopped being a slide-deck idea in the past year. OpenAI and Stripe, Google, Visa, Mastercard and the major commerce platforms all shipped working pieces. Gartner predicts that by twenty twenty-eight, ninety percent of B2B buying will be handled through AI agents. That's more than fifteen trillion dollars of spend. Treat a forecast like that as a direction rather than a budget line. B2B sellers aren't first in line. They are in the line.
What an agent does when it shops
A buyer, or a purchasing team, tells an assistant what it needs. The agent reads product data, builds a shortlist and compares. This agent works for the buyer, and it may never look at your pages the way a person does. Stage one is discovery. The agent reads product feeds, structured data and pages to build a shortlist. Most of the activity is here today, and incomplete product data quietly drops you out of the comparison. Stage two is delegated checkout. The agent completes the purchase inside the AI app, through a checkout API and a payment token the buyer approved. Stage three is autonomous buying within set rules, with no person present, such as replenishment or machine-to-machine payments. Google's AP2 version zero point two and Mastercard's Agent Pay for Machines are built for this, and it's still early.
ACP and UCP checkout protocols
Several open standards shipped in twenty twenty-five and twenty twenty-six. They overlap more than they compete. Two of them handle checkout. The Agentic Commerce Protocol, or ACP, came from OpenAI and Stripe in September twenty twenty-five, alongside Instant Checkout in ChatGPT. It's open source under Apache two point oh, and still labeled beta. Its April twenty twenty-six release added carts, a product feed API and an MCP binding. In March, OpenAI said the first Instant Checkout wasn't flexible enough, and moved ChatGPT's focus to product discovery, with merchants using their own checkout. The Universal Commerce Protocol, or UCP, was announced by Google in January twenty twenty-six, built with Shopify, Etsy, Wayfair, Target and Walmart. It runs checkout in Google's AI Mode and Gemini for eligible US retailers. In April, Amazon, Meta, Microsoft, Salesforce and Stripe joined its Tech Council.
AP2, MCP and card networks
The Agent Payments Protocol, AP2, is Google's payments layer. It launched in September twenty twenty-five with more than sixty partners. It keeps signed records, called mandates, of what the buyer approved. In April twenty twenty-six, Google handed it to the FIDO Alliance. The Model Context Protocol, MCP, is Anthropic's open standard for connecting AI models to tools and data. It was released in November twenty twenty-four and is now part of the Linux Foundation's Agentic AI Foundation. ACP and UCP can both run over MCP, and commerce platforms now ship storefront MCP servers. Then there are the card networks. Visa's Trusted Agent Protocol, built with Cloudflare, and Mastercard Agent Pay let agents pay with tokenized cards, and help merchants tell a trusted agent from a scraper.
Which protocol to bet on
If you're wondering which one to bet on, you don't have to. Your commerce platform implements the protocols, and the platforms are adopting more than one. Your job is the data and the rules underneath them.
Where the commerce platforms stand
So where do the commerce platforms stand? This is as of September twenty twenty-six, from each vendor's own announcements. BigCommerce added Stripe's Agentic Commerce Suite, endorsed UCP, and announced agent-ready product feeds to ChatGPT, Gemini, Copilot and Perplexity through Feedonomics. Its Storefront MCP server is in beta for guest shopping, and the B2B server is listed as coming soon. Shopify's Agentic Storefronts sell into ChatGPT, Copilot, Google AI Mode and Gemini, and every store has a Storefront MCP server. Salesforce Commerce Cloud announced support for ACP and UCP, and made its B2B Buyer Agent generally available in June twenty twenty-six. Adobe Commerce committed to UCP and ACP, and introduced a Commerce MCP server covering catalog, cart, pricing, inventory, checkout and orders.
Why B2B is harder
Consumer agent checkout assumes one buyer, one list price and one card. B2B buying runs on contract pricing, purchase orders, payment terms, approval chains, and punchout from procurement systems. And the protocols have barely started on that.
What the protocols cover for B2B
UCP's August twenty twenty-six release added fractional quantities, payment terms and split payments. Its release notes don't mention purchase orders, contract pricing or approvals. In ACP, the only change labeled B2B so far is support for decimal quantities. And we found no cXML or punchout support in either protocol. The vendors agree. Adobe names complex pricing, multi-step approvals and procurement workflows as the hard part, and BigCommerce says authenticated shopping with personalized pricing comes next.
Buyers won't wait
Buyers won't wait for the standards to catch up. Forrester predicted that in twenty twenty-six, one in five B2B sellers would have to negotiate quotes with buyers' AI agents. Those agents will read whatever you publish. So the practical question is whether what you publish is right.
Fix your product data first
None of this requires picking a protocol. It requires data an agent can trust, and rules an agent can follow. Start with product data. ChatGPT's product feed spec requires an item ID, title, description, URL, brand, image, price and availability. Google added Merchant Center attributes such as product FAQs and compatible accessories for conversational search. Missing specifications leave you out of the comparison.
What a B2B seller should do now
Keep price and stock true to the ERP. An agent that quotes a stale price, or promises stock you don't have, creates a dispute instead of a sale. The ERP stays the source of truth. Publish your terms where machines can read them. Minimum order quantities, lead times, shipping, returns and payment terms belong on crawlable pages and in structured data, and not only in PDFs or behind a login. Expose clean APIs. On BigCommerce that means the Storefront MCP server and B2B Edition APIs. On a custom stack, it means an MCP endpoint with scoped access. Decide what an agent may do alone. Which orders can it place on a buyer's account, up to what value, and what needs a person to approve? Write it down before the first agent order arrives. And let the AI crawlers in, and watch them. Keep an llms.txt file current, and check your server logs. On our own site, requests identifying as ChatGPT-User fetched pages on every one of the thirty-two days to September fourteenth.
Wrap-up
To sum up: your platform implements the protocols, and they change month to month. What's on you is complete product data, prices and stock that match the ERP, clean APIs, and clear rules for what an agent may do. ProjectThunder builds B2B storefronts, product data and ERP integrations that agents can read and trust. The full write-up, with links to the related guides, is on projectthunder.com. Thanks for watching.
Narration is an AI voice (Azure). The facts come from the linked write-up.


