What's already true
Personal AI assistants that act, not just answer, now exist as products. Meta's Muse has its own browser, opens pages and fills in forms. Instinct gives each user's agent its own email address; it can contact businesses, ask about availability and create accounts. OpenAI's ChatGPT agent and Anthropic's Claude in Chrome do multi-step work in a browser.
None of them is built for procurement, and none of this shows anyone buying building products through an agent. What it shows is the shape of the tool: software that remembers what you want, does the research, and contacts businesses for you.
Business buyers already use AI for supplier research. In surveys of B2B buyers across industries, Gartner reports that 45% used generative AI in a recent purchase and 69% checked its output with a sales rep. Forrester reports almost all business buyers now use AI somewhere in buying. Those surveys are mostly about technology purchases, not construction.
In construction, the picture is earlier and mixed. The AIA found in 2025 that 6% of architecture professionals use AI regularly, while 74% see value in it for product research. A 2026 Farnsworth Group and Venveo guide reports nearly 40% of architects use AI tools for product research. Those numbers use different definitions a year apart. Both say something is starting. Neither says it has arrived.
What we expect, and why
This part is our hypothesis, not a finding.
People don't keep a wall between useful personal software and useful work software. In Microsoft's 2024 Work Trend Index, 78% of people using AI at work were bringing their own tools. Dropbox spread the same way: people used it for themselves, then brought it to work, and companies followed. That doesn't mean AI agents will follow Dropbox's path. It means the bottom-up pattern is familiar.
Construction may get there through its smallest firms first. A two-person GC or a remodeler has no IT department deciding which assistant is allowed. If the assistant that books their flights can also find a supplier for a job, they'll ask it to.
So some product research will start to look like this. The buyer describes what the project needs. The agent researches products and suppliers, compares them against the requirements, checks fit and location, and contacts a supplier with a structured request. The buyer reviews what comes back.
A hypothesis: agents arrive at work the way personal tools always have, from the bottom up.
Why building products are hard for AI
A consumer product has a SKU, a price and a photo. A building product usually has a family with options, where the performance depends on the configuration: a window's U-factor and design pressure vary with size and glazing. It has values tied to a test standard: NFRC for windows, ASTM methods for air barriers, UL for fire ratings. It has approvals with an issuer and a date, like an ICC-ES evaluation report, a UL listing, a Florida Product Approval or a Miami-Dade NOA, and some of them expire. It has documents that hold the real detail: spec sheets, submittals, installation guides, CAD and BIM files. And it has a channel: a manufacturer, a rep agency, one or two levels of distributor, a dealer, and sometimes an authorized installer.
Each of those lives somewhere different, and they don't always agree. A distributor's page lists last year's series. A directory files a supply-only company under installers. The territory map is in a rep's head.
What we saw when we asked
On 2026-10-06 we ran eight building-product questions through a web search and its AI summary. This is a small sample, not a study.
Spec questions worked best. Queries for an ASTM air-barrier standard, a 90-minute fire-rated glazing product and an NFRC U-factor surfaced manufacturer spec pages, spec platforms (Sweets, ARCAT, NBS Source), trade press and PDFs.
“Where do I get it” questions worked worst. Asked who distributes a well-known storefront brand near Cleveland, the summary couldn't name a distributor and told us to call the manufacturer. “Alternatives to” questions went to forums and how-to sites, not to manufacturers' own comparison pages. Rep agencies didn't appear at all.
That matches what a buyer's agent would need most and find least: who can supply this product near this job, and who quotes it.
AI finds the spec. It can't find who sells it near the job.
What has to be true
Take the request again: three window systems, one spec, a supplier serving Northeast Ohio, enough to quote. For a company's products to make the list:
Discover. AI knows the company exists, what it sells, and where.
Understand. The product pages say what each family is, what it's made of and where it's used, in text, with the documents linked.
Qualify. Each important value carries its test standard and date, so it's read rather than guessed.
Contact. It's published who quotes that region: the manufacturer, a rep, a distributor or a dealer. An agent shouldn't have to assume the maker sells direct.
Act. A structured request reaches a person with what they need. It's a request, not an order: no price, no stock promise, no delivery date.
What this means for manufacturers, suppliers, distributors and dealers
The work isn't “optimizing for ChatGPT”. It's keeping one accurate record of the company and its products, with a source and a date on every fact, publishing it the same way everywhere, checking what AI says against it, and giving an AI assistant a controlled way to ask for a quote.
At Lehvel we call the company side Business Truth and the product side Product Truth. The record also says what's private: some supplier relationships shouldn't be published, and an AI assistant should never learn them from you.
