There are now more than a hundred startups marketing AI products to general contractors, and most of their sales decks are functionally interchangeable. They all promise faster project discovery, AI-powered matching, and a smarter pipeline. A meaningful share of them are lead lists wearing a chatbot.
The good news is the diagnostic is fast. There's a ninety-second test you can run on any of these products that almost always exposes whether the underlying work is real construction-intelligence engineering — or whether the product is wrapping public-data scrapes with a search bar and a generative summary layer.
Here's the test. If you're evaluating any of these products, run it.
The 90-second test
Ask the salesperson to show you a single project, in your market, that they consider a representative win. Pick one. Then ask:
- 01 ·When was the plat filed on this parcel? Show me the document.
- 02 ·Who's the architect of record on the most recent plan review? Show me the file.
- 03 ·What's the ownership LLC, when was it formed, and who are the managers per the Texas SOS filing? Show me.
- 04 ·What's the inspections record on any active permit on this parcel? Show me.
Four questions. Each one corresponds to a record that's publicly available, costs nothing to surface, and is the kind of thing a real construction-intelligence product should be able to produce in real time.
Most products in this category cannot answer all four. The ones that can are doing the real work of ingesting and resolving public records; the ones that can't are wrapping a thinner upstream feed. When a product fails the test, the salesperson typically responds by redirecting — to scoring methodology, to AI architecture, to CRM integrations, to a follow-up call "to dig into the data." The redirect is the tell.
What "real" AI in construction actually does
I want to be careful here because the phrase "AI for construction" has been so devalued by marketing copy that it almost means nothing. So let me say what it means concretely, from the perspective of a contractor trying to evaluate a vendor.
Real AI in construction intelligence is doing three things: ingestion (pulling from disparate public sources at scale), resolution (joining records about the same project across plats, permits, ownership, and timelines), and scoring (predicting which projects are worth your team's time given your specific business). Each of those is a real engineering problem. None of them is what most products in this category are doing.
What most products are doing instead: buying a project-leads feed from a single source (often the same upstream vendor everyone else uses), wrapping it in a search UI, and adding a chatbot that can summarize records you could read yourself. That's not AI. That's a search interface with a generative add-on.
The three failure modes
When a product fails the permit test, it usually fails for one of three reasons.
The single-source problem. The product is built on top of one upstream data feed — usually a national permit aggregator that covers building permits but nothing earlier. They have permits but not plats. They have addresses but not ownership LLCs. They have project descriptions but not architect names. The result is a product that confirms what's already happening, without any of the early-stage records that would make pursuit valuable.
The unresolved-records problem. The product has multiple data sources but hasn't done the work of joining them. A plat, a permit, and a parcel record on the same piece of land show up as three different rows. The user has to do the resolution work themselves, which means the product is functionally a glorified spreadsheet with a search bar.
The contact-enrichment-disguised-as-intelligence problem. The product is really a contact-data product. They have phone numbers and emails for thousands of construction industry contacts. They wrap that in language about "AI-powered project intelligence" because that's what the market wants, but the actual offering is a contact database with a search filter. This is just a lead list with a friendlier UI, and the failure mode is the same.
What to ask in a demo
Beyond the permit test itself, a few specific questions that tend to expose the real shape of a vendor:
- What's the median lag between a public-record event and your platform reflecting it? Real-time-ish answers (under 48 hours) are good. "Updated weekly" is suspect. "We refresh the dataset monthly" is a lead list.
- How do you handle parcel records vs. project records? If they don't understand the question, the product is not doing record resolution.
- Show me the source for the contractor-of-record field on this permit. If they can show you the public document, great. If they can only show you the field in their UI, they're trusting their upstream vendor without verification.
- What's your coverage in [my specific metro]? Drill on metros, not on "national." "National" almost always means the coverage in your specific city is thin.
The honest tradeoff
Two genuine tradeoffs are worth acknowledging here.
First: products that do real ingestion, resolution, and scoring are more expensive and slower to scale across metros than wrapped lead lists. A product that covers 80 metros nationally is almost certainly thinner per metro than one that covers 8 metros deeply. I've written about this asymmetry before — the depth/breadth tradeoff is real, and not every contractor needs deep coverage of every metro they touch.
Second: even a great construction intelligence product is not going to win bids for you. It will surface projects, time the outreach, and give you context. The conversations, the relationships, the pricing, the execution — those are still you. AI in this category is genuinely useful, but it's not a replacement for sales discipline. Vendors who talk about "AI-powered conversion" or "automated outreach" without acknowledging the human side of the work are usually selling something that's going to disappoint a sales team within ninety days.
Run the permit test on every vendor you evaluate. If they pass, you've narrowed the field meaningfully. If they don't, you've saved yourself a 90-day pilot and a few thousand dollars in subscription fees you would have written off.