How AI Specification Changed B2B Buying

What actually changed?

The change is not that buyers use AI. It is where in the sequence they use it.

For thirty years the industrial buying journey started with a search, a distributor site, a catalog or a rep, and ended with a shortlist. The supplier's job was to be present at the start — through SEO, through franchise distribution, through field application engineers who sat with designers. Presence early produced consideration late.

That sequence has inverted. The buyer now starts with an assistant, arrives at a shortlist, and only then touches a supplier surface — usually to verify a decision that has already narrowed. Forrester's January 2026 analysis found 94% of business buyers using AI somewhere in their buying process, up from 89% a year earlier, and — more significant than the headline — that twice as many buyers named generative AI or conversational search as a more meaningful or important source of information than any other, outranking vendor websites, product experts and salespeople.

Read that carefully. It is not that AI is one channel among several. It is that AI has displaced the channels the industry spent decades building.

Where does the evaluation actually happen now?

Three datasets triangulate the same shift from different angles.

The Adobe numbers are retail, and industrial buying is not retail. But the mechanism they expose is channel-agnostic: when an assistant does the filtering, the visit that follows is a decision, not a browse. Adobe also found those visitors spent 48% longer on site and viewed 13% more pages. Fewer, later, better-qualified — and entirely dependent on the assistant having chosen you.

The scale underneath this is not marginal. OpenAI reported more than 900 million weekly active ChatGPT users in February 2026. Whatever share of those are engineers checking a part number, it is a larger population than any distributor's field organisation has ever reached.

Why does this hit specification harder than it hits marketing?

Because specification is where industrial margin is actually made, and specification is unusually easy for a machine to influence.

A marketing funnel degrades gracefully when a channel weakens. A specification does not degrade — it resolves. A part is either in the bill of materials or it is not, and once it is in, it is expensive to remove. Design-in cycles of 12 to 36 months mean the decision and the revenue are separated by years.

The traditional mechanism for capturing that value is design registration: a supplier rewards the distributor whose field engineer got the part designed into a customer's product, typically with additional margin, price protection or sole-sourcing on that programme. The whole apparatus — franchise agreements, demand creation targets, FAE headcount — exists to answer one question: who caused this design-in?

Assistant-led specification breaks the question rather than the answer. If a design engineer describes a requirement to an assistant, receives three candidate parts with parametrics, and registers a design against one of them, no field engineer created that demand. The data did. And nobody has agreed how to credit data.

This was already a strained mechanism. Design registration has long struggled when a customer outsources production to a contract manufacturer, because the design happens in one place and the purchase order appears somewhere else entirely — the registration survives on paper while the commercial logic frays. Machine-assembled shortlists apply the same stress at the front of the process instead of the back. The distributor that invested in the data is not necessarily the one holding the relationship, and the manufacturer whose parametrics were legible may never have been called on at all.

What determines whether your part appears in the answer?

It is worth being concrete about the mechanism, because most commercial teams reason about it as though it were a ranking.

An assistant answering a specification question does roughly four things: it interprets the requirement into constraints, retrieves candidate records, filters them against the constraints, and then justifies the survivors. A part can be eliminated at any of the four, and only the last resembles anything a marketing team has previously optimised.

The second row is the quiet killer. An assistant asked for an AEC-Q100 Grade 1 part will not include a component whose qualification status it cannot verify. It does not report the exclusion, and it does not hedge. Absence of data is treated as absence of the property — which is the correct behaviour for a cautious tool and a catastrophic outcome for a supplier whose part actually qualifies.

Who captures the value when the shortlist is assembled by a machine?

The pattern is consistent: leverage moves from who you know to what a machine can verify about your parts. That is an uncomfortable finding for organisations whose competitive moat is a field organisation, and a genuine opening for anyone with good data and no field organisation at all.

Is this a rupture, or just a new channel?

The honest answer is that the discovery shift is well evidenced and the transaction shift is not.

Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. That prediction has aged well: OpenAI withdrew its own Instant Checkout in March 2026 after roughly a dozen Shopify merchants went live, and Deloitte reports fewer than a quarter of B2B suppliers use agentic AI technologies at all.

Industrial buying has additional friction that retail does not. Contract pricing is negotiated and confidential. Approvals involve the 13 internal stakeholders and 9 external influencers Forrester counts, with procurement acting as decision-maker in 53% of cycles. Compliance documentation — RoHS, REACH, conflict minerals, UKCA and CE declarations — has to be attached to the decision, not asserted about it.

None of that stops an assistant from writing the shortlist. It stops an assistant from placing the order. Those are different claims, and conflating them is the most common error in this market.

What should a manufacturer or distributor do about it?

1. 01Find out what the assistants currently say about your parts. Ask five factual questions with known answers about five real part numbers, across three assistants. Most teams discover a specific, correctable error inside twenty minutes.
2. 02Fetch your own catalog pages with a plain non-browser client. When Partsgraph audited 984 distributor and manufacturer domains worldwide in August 2026, 38% served no readable catalog page to a standard non-browser client, and the median AI-visibility score was 50 out of 100. Client-side rendering is the single most common cause.
3. 03Make one record canonical. Parametrics scattered across a PIM, a PDF datasheet and an ERP will produce three different answers. An assistant will quote whichever it found, without telling anyone which.
4. 04Publish compliance and lifecycle status as data, not as documents. Obsolescence, last-time-buy dates and regional approvals are the fields most likely to be wrong in an assistant's answer and most likely to be commercially fatal when they are.
5. 05Give agents a direct route. Microchip published an official Model Context Protocol server in November 2025 exposing verified product specifications, datasheets, inventory, pricing and lead times to AI clients. It remains a conspicuously short list — in the same August 2026 audit, not one of the 984 domains advertised an MCP endpoint.
6. 06Rewrite the demand-creation rules before your channel does. If design registration cannot credit machine-originated specification, distributors will stop investing in the data that produces it, and the manufacturer inherits the problem.

The part that compounds

There is no plausible future in which buyers go back to opening twelve tabs. Once an assistant answers a specification question correctly, it becomes the default source for that class of question, and defaults are self-reinforcing: the answer that gets cited gets reinforced, and the supplier that is legible gets cited.

The lag makes this worse, not better. A specification won in 2026 arrives as revenue in 2028. Which means the companies discovering in 2028 that their design-in rate has quietly declined will be looking at data quality decisions made two years earlier, by nobody in particular, in a system nobody owned.

If you want to know what assistants currently say about your catalog, the free grader at [/audit](/audit) checks how a machine actually reads it.