Agentic commerce is a B2B story, and the suppliers are behind.
The job worth paying for is not building your company an AI agent. It is making your company transactable by whatever agent shows up.
That is the whole point of view, and almost every agentic commerce pitch I have seen gets it backwards. They want to sell a manufacturer or a distributor an agent of its own: a chatbot on the site, an assistant in the portal, something with a name. Meanwhile the buyer already has an agent, it is not yours, and it is deciding whether your company is easy or hard to buy from.
The buyers moved first
Deloitte Digital surveyed more than a thousand US B2B suppliers and buyers and published the results in February 2026. Nearly 40% of B2B buyers said they already use agentic AI in purchasing, for evaluating products, configuring orders, reviewing contracts and benchmarking prices. On the supplier side, 24% said they use agents in the sales process. Most of the rest said they plan to.
The same survey found a perception gap that should worry anyone who runs a sales organization. Seventy-two percent of suppliers described their own sales process as mostly or highly automated. Only 47 percent of their buyers agreed. Buyers were several times more likely than suppliers to call the process manual.
What an agent actually does to you
Here is the part that most of the excitement skips. A procurement agent optimizes against you, not for you.
That is not cynicism. It is the job description. The buyer's agent is built to find the right part, at the right spec, at the best available price and terms, from a supplier that can actually fulfill. It does not care about your relationship with the buyer. It does not remember that you carried the account through 2009. It reads what is machine-readable, compares it, and moves on.
So when an agency offers to build you an agent, ask what problem it solves. Your buyer does not need your assistant. Your buyer needs your business to be legible to theirs.
What transactable means
Being transactable by whatever agent shows up is a concrete list, and none of it is glamorous.
Account and catalog modeling. Company accounts, locations, buyer roles, and the catalog each one is allowed to see, structured so a system can ask "what does this location pay for this SKU" and get an answer without a human.
Product data with real depth. Not a title and a picture. Specifications, units of measure, case packs, materials, certifications, cross references, fitment. The data an engineer would need to be sure it is the right part, in a form a machine can read.
Machine-usable quoting. A way to ask for a price on a configured or non-stock item and get a structured answer back, rather than an email that starts "let me check with Gary."
ERP write-back. Orders, inventory, customers, pricing and fulfillment moving in both directions between the storefront and the system of record, so the number the agent sees is the number that is true.
Integration with your buyers' systems. Punchout, EDI, whatever their procurement stack speaks. The agent is going to live in that stack, not on your site.
You will notice that none of these are AI features. They are data and integration work, and they are the same work that has always separated a B2B storefront that gets used from one that does not. The agents did not change what the work is. They changed how expensive it is to skip.
The defensible position
If the agent is optimizing against you on price and spec, where is the defensible ground?
It is not being cheapest. Somebody is always cheaper, and the agent will find them.
The defensible position is being connected and hard to substitute. Connected, meaning your systems already talk to the buyer's, so switching away from you costs them integration work. Hard to substitute, meaning your product data answers questions that a commodity listing cannot.
This is why cross reference, interchange and fitment data matter more in the agentic world than anywhere else. Most distributors already own this data. It lives in the heads of the counter staff, in a spreadsheet somebody built in 2014, in the ERP under a field nobody exposes. It says which part replaces which, what fits what, what is equivalent to the thing the buyer typed in wrong.
A commodity marketplace does not have that data and cannot easily get it. Your competitors across town may not have it either. When a buyer's agent asks "what replaces this discontinued bearing in this housing," the supplier whose catalog can answer wins the order, and the one whose catalog says "call for details" does not get called.
That data is the moat. Almost nobody is using it.
What to do first
Do not buy an agent. Do not let anyone sell you one until the following is true.
Your product data would survive being read by a machine that does not know your industry. Your customer-specific pricing lives somewhere a storefront can read it in real time, whether that is the ERP or the storefront itself. Your ERP and your storefront agree on inventory and price without a human reconciling them. And your biggest customers can send you an order from their procurement system without re-keying it.
If most of that is not true yet, the agentic conversation is premature, and any agent you build will be confidently wrong about your own stock and prices. The same messy data that keeps a B2B storefront from working keeps an AI from working. It is one problem.
This is the work a B2B build with ERP integration exists to do: the storefront, the company accounts and catalogs, and the integration that keeps the storefront honest against the system of record, with the commerce logic decisions that a connector alone will not make for you. Integration is scoped into the build around your ERP, connector availability and the objects that need to move, and every build starts with Discovery, which tells you whether your data is ready for any of this.
The buyers already have their agents. The question for a supplier is not how to build one. It is whether, when one shows up, your business answers.