Picture a Tuesday morning sometime in the next two years. A purchasing agent working for a customer you have never met sends a structured request to forty suppliers at once: a part described by material, tolerance, finish, quantity, required certifications, and a delivery window. Nobody calls. Nobody opens your website, reads about your fifty years in business, or meets your sales rep. Within minutes the agent has a ranked comparison, and the suppliers who could not answer in a form it could read are not on the list at all.
Not the day a robot buys a pallet of your product, but the day a machine decides whether you are even considered.
That is the moment this article is about. Not the day a robot buys a pallet of your product, but the day a machine decides whether you are even considered.
The forecast, and why you do not need it to be right
In its Strategic Predictions for 2026, published in November 2025, Gartner predicted that by 2028 some 90 percent of B2B buying will be intermediated by AI agents, pushing more than $15 trillion of B2B spend through agent exchanges. Treat that as what it is, a prediction with a date on it, and treat the number with the same caution you would give any analyst forecast. We keep our own calls in a dated, scored ledger at our predictions page for exactly this reason.
You do not need the 90 percent to be right. You need only the direction to be right, and the direction is already visible: buyers are putting software between themselves and their suppliers. The software does not read brochures. It reads fields, and it screens on what it can verify. That changes what counts as being easy to buy from.
First, sort your lines
Not every manufacturer faces the same exposure, and the first piece of work is deciding which of your lines are in the line of fire. Our 2026 AI Dividend Map places industrial manufacturing in the Rising Bar shape, with a specific warning attached: commodity exposed. Capacity and capital keep any one manufacturer from taking the whole market, so you need not fear a monopolist. What commodity manufacturers should fear is their customers. When procurement agents source on spec and price alone, commodity lines become price-takers, because the gate that once kept buyers from shopping around opens from the demand side. We describe those gates in Shapes of the AI Dividend. Differentiated and custom work stays more sheltered, and for those lines operational AI is simply the cost of quoting competitively.
Here is a practical test for each product family. If a buyer could describe the part in a specification and three other shops could make it to that specification, an agent will treat it as a commodity. If the buyer needs a conversation with an engineer before anyone can price it, an agent cannot finish the job alone. Most mid-market manufacturers have both kinds of work, and the right moves differ. For the commodity lines, readiness decides whether you make the shortlist. For the custom lines, it decides whether you are reachable when the first machine-run screen happens.
Make the part describable to a machine
Agents compare attributes. If your specifications live in PDF catalogs, in a salesperson’s memory, and in an ERP full of free-text descriptions, the agent cannot compare you with anyone, so it will not try for long. The fix is structured product data, and you do not have to invent the structure. Classification standards already exist for this purpose.
ECLASS describes itself as a global reference data standard for the classification and unambiguous description of products and services, and says it follows ISO 8000 for data quality. ETIM calls itself the international classification standard for technical products: it assigns each product to a class, gives each class defined characteristics, and gives each characteristic fixed values or ranges so that IT systems can compare, search, and filter automatically. Its published examples lean toward electrical and building products, so check which standard your own channel and customers already use. In procurement systems you will also run into UNSPSC for spend categories and GTIN for product identifiers. You do not need all of them. You need the one your buyers’ systems already speak, populated consistently.
Start where the quotes are, not with the whole catalog. List the part families that generate most of your quote requests, and for each one fix the unit conventions, the tolerance fields, the material grades, and the standard identifiers, so that the same fact is written the same way every time. A description that says “SS bolt, fine thread” will lose to a record with a material grade, a thread pitch, and a length in millimeters, even if your bolt is better.
Make your claims checkable
A human buyer will forgive a vague answer about certifications and call you back. An agent screening forty suppliers will not. It filters first on hard requirements: quality management certification, industry-specific approvals, material test reports, restricted-substance compliance, and registration requirements. If those exist as scanned PDFs on a shared drive, with expiration dates known only to the quality manager, then to a machine you are unqualified, and you will never learn it happened. This is our reading of how screening will work, not a measured result, but it follows from how a filter works. A missing field fails quietly.
A buyer’s system that gets a promise you cannot keep can be expected to log the miss, and a logged miss follows you into the next request.
Build a certificate register your systems can read: each credential, its scope, its issuing body, and its expiration date, kept current and published where an authorized buyer can retrieve it. Do the same for capacity and lead time. Do not publish a single optimistic number. Publish the lead time you can actually meet at current load, drawn from the schedule, and update it. A buyer’s system that gets a promise you cannot keep can be expected to log the miss, and a logged miss follows you into the next request.
Quote at the speed the buyer expects
Think about how a quote gets produced in your shop today. A request arrives by email, an estimator pulls drawings, checks material prices, asks scheduling about capacity, builds a spreadsheet, and replies in a day or a week. A human buyer tolerates that. An agent that has sent the same request to forty suppliers works to a window, and the answers that arrive inside the window are the ones that get compared. We expect quote turnaround to become a ranking input, for the simple reason that it is the easiest thing for a buyer’s software to measure.
You do not have to automate everything, and you should not try. Divide quotes into three groups. Standard parts, within capacity, above a margin floor you have defined, can be priced by rule: write the price breaks and the lead-time logic down, because in most shops those rules live only in the estimator’s head. Anything unusual goes to the estimator with a clock on it, so a human answer arrives while it still matters. Anything above a dollar threshold you set goes to a person for approval before it leaves the building. Machine-speed quoting makes machine-speed mistakes, and the margin floor and the approval threshold are what stop one bad rule from repeating across a hundred quotes.
Open a berth for the agent
A contact form is built for a person. An agent needs something it can call. We cover this in detail in Your Next Customer Won’t Be Human, which describes an agent dock with five parts, one of which is the berth: an externally reachable MCP server exposing a narrow set of tools. For a manufacturer, that narrow set can be small. A tool that checks whether a part number is orderable and in what quantity. A tool that returns price breaks and a realistic lead time for a specification. A tool that submits a request for quote into your estimator’s queue. A tool that reports the status of an existing order.
Start read-only, with the first two, on one product family. Keep the tool list short on purpose, authenticate the buyers you know, and log every call. Also remember that agents will not only arrive at your front door. They will read your distributors’ copies of your data, supplier directories, and marketplace listings, so the clean data you build for your own berth should flow to every channel that carries your products. Stale data in a distributor’s catalog will be read as yours.
Decide what the machine does not get
Opening a door to buyers’ agents opens it for competitors’ agents too. A public price tool lets anyone map your price curve in an afternoon, and a drawing library with no controls hands your engineering to people who never signed an agreement. Decide in advance which prices are public, which require a buyer account, and which only appear after an NDA. Put rate limits on the tools. Treat drawings and CAD files as controlled documents, not as website content.
This is the same logic as clearly marked lanes: speed is safe when the boundaries are written down first.
Run the mystery shopper test
Before you spend anything, take your best-selling part number and your three most important capabilities, open a general-purpose AI assistant, and ask it to find qualified suppliers. Then ask it what it knows about you. See whether you appear, whether the facts are right, and whether it can tell what you certify, what you can deliver, and how to request a quote. Most manufacturers will find this humbling, which is useful. It is the cheapest audit you will do all year.
Then sequence the work over a quarter. In the first month, sort your lines into commodity and custom, and audit the data quality for the part families that drive your quotes. In the second month, fix the worst gaps, publish the certificate register, and write down the pricing and approval rules. In the third month, stand up a read-only berth for one family and test it with an agent of your own. If you want to see where the data foundation stands across mid-market manufacturers, our Analytics Maturity in Manufacturing survey is a useful comparison point.
What readiness will not do
Being readable does not save a commodity line. It gets you on the shortlist, where an agent will compare you on price and specification and nothing else, which can accelerate the competition you were hoping to avoid. The durable defense is the one the Dividend Map points to: move lines out of the commodity category, through engineering depth, service, certifications that few shops hold, or faster turnaround on hard parts, so that a specification alone cannot describe what a buyer is buying.
The durable defense is the one the Dividend Map points to: move lines out of the commodity category, through engineering depth, service, certifications that few shops hold, or faster turnaround on hard parts, so that a specification alone cannot describe what a buyer is buying.
It is also worth saying plainly that many B2B orders, custom work especially, will still involve people, and the Gartner date may prove early or late. The work above is not wasted either way. Clean product data speeds your own estimators, improves what distributors can sell, and shortens onboarding for every new buyer, human or not.
If you would like help sorting your lines and testing how a machine sees your company today, an independent assessment is a good place to start, and our manufacturing IT and AI consulting practice works on exactly this kind of readiness. Either way, do the first step this week. Ask an AI assistant who makes your best-selling part, and see whether your name comes up.
