Analysis. Two things were published this month that are hard to hold in your head at the same time.
On July 15, McKinsey released a study on AI in architecture, engineering and construction. As reported by Construction Dive, it assessed 150 workflows across 25 AEC domains and concluded that AI could automate 39% of nonphysical work in construction, and 50% of nonphysical work in architecture and engineering.
Nine days earlier, the Daily Commercial News reported survey findings from a global DEWALT study of construction professionals across Canada, the US, the UK, France, Brazil and Mexico. The share who use AI in their day-to-day work: 9%. Another 39% are piloting, 35% are researching, and 16% are interested but have not started.
Both can be true. The interesting question is what the distance between them actually measures.
The Gap Is Not a Forecasting Error
The easy read is that McKinsey is wrong, or early, and the 9% is the real world catching up eventually. That read is too comfortable, because the two numbers are measuring different things.
McKinsey’s 39% is a statement about task structure — how much of the nonphysical work in construction is, in principle, automatable by current technique. It is an engineering claim about the shape of the job. The 9% is a statement about behavior — how many people opened the tool this morning. A technology can be capable of automating 39% of your paperwork and used by almost nobody, and neither fact contradicts the other. The gap is the adoption problem, stated numerically.
What makes the gap unusually informative this year is its direction of travel, which is worse than the headline suggests. The Scius Advisory AI in Construction Report 2026, also cited in the DCN piece, tracked stated intent to adopt AI software from 8% in Q2 2024 to 9.6% in Q2 2025 — 1.6 points in a year. Hardware intent moved from 2.8% to 3.6%. Those are not the slopes of a technology about to inflect. They are the slopes of a technology being watched.
Where the Money Went Instead
The capital markets did not wait for the slope. Cemex Ventures’ Q2 and H1 2026 investment report, published July 8, found that roughly 70% of all contech transactions in the period involved AI-enabled startups, and drew the obvious conclusion: “AI has stopped being a differentiator and started being a baseline expectation.”
The underlying volumes are more sobering than that sounds. Q2 2026 saw $1.031 billion across 84 deals — capital up 30% year over year, but deal count down 8%. Across the half, disclosed investment was $2.884 billion across 153 deals, down 19% in capital and 20% in deal count against H1 2025. Cemex notes the true figure is likely nearer $4.5 billion once three undisclosed acquisitions are counted.
Read together: fewer companies are getting funded, the ones that do are getting more, and nearly all of them are selling AI into an industry where 9% of practitioners use AI daily. That is a market where the investment thesis has run ahead of the usage data — which is not automatically irrational, but does mean the returns depend on adoption accelerating rather than continuing at 1.6 points a year.
The Barriers Are Not the Ones Vendors Are Solving
The DEWALT respondents were asked what is stopping them. The answers cluster: 23% cited a lack of training and skill gaps, 23% cited accuracy and trust, and 20% cited privacy, cybersecurity and IP risk. Notably, 45% reported no concerns at all about adopting AI, against 35% who did — so this is not a wall of hostility. It is closer to indifference plus unresolved doubt.
Almost none of that is a model-capability problem, and almost all of the sector’s engineering effort is going into model capability. Training gaps are solved by deployment services. Trust is solved by showing work and by being wrong in visible, correctable ways. IP risk is solved by contract terms and data residency. These are go-to-market and product-surface problems, and they are unglamorous relative to shipping another agent.
There is a reading of the vendor behavior that fits. Procore’s Digital Coworker packaging in late July shipped Control Tower — an admin console for seeing which agents consume which credits — and previewed Skills, which lets a contractor teach agents its own standards from its own documents. Neither is a capability advance. Both are trust-and-control features. The incumbent with the most adoption data built the things the survey says are missing.
The “Two Camps” Framing Is the Useful Part
The most quotable line in the McKinsey work is not a number. The report divides AI users into those automating core tasks and those using AI as, in Construction Dive’s paraphrase, a superficial productivity tool.
That distinction does real work as a filter. It maps onto the difference between a contractor whose estimating process now begins with a model and one whose estimators occasionally paste a spec into a chatbot. Both would answer “yes” to a survey question about using AI. Only one of them is on the curve McKinsey is describing. Which means the 9% figure is, if anything, generous about depth — daily use includes a great deal of the second category.
It also explains McKinsey’s sequencing, which is more conservative than most vendor roadmaps. Near-term, within 18 months: bid/no-bid analysis, estimating, proposal drafting. Medium-term, 18 months to four years: work built on proprietary data — RFIs, drawings, specs, close-out reports. Long-term, four-plus years: autonomous construction equipment and transportation coordination. The industry’s most-funded category, jobsite robotics, is in the four-plus-year bucket.
What This Means If You Are Buying
Three things follow, and they are the kind that survive whichever number turns out to be closer.
The first is that the near-term value is in the boring end. Estimating, bid/no-bid and proposal drafting are where McKinsey puts the 18-month payoff, and it is where the trade-vertical estimating startups have quietly concentrated — Guthrie AI in glazing, Bidflow in electrical, XBuild in roofing. That is unfashionable relative to robotics and better matched to the timeline.
The second is that proprietary data is the medium-term moat, which is the same conclusion the platforms reached with their chequebooks — Trimble buying Document Crunch, Procore buying DroneDeploy, Suffolk arguing its 293 terabytes are the asset.
The third is the one worth acting on now: if 23% of your people cite training and 23% cite trust, then the binding constraint on your AI programme is not which vendor you pick. Deployment, not procurement, is where that 39% actually gets claimed.
A note on sourcing: the 39%, 50% and 150-workflows figures are as reported by Construction Dive from McKinsey’s July 15 report; McKinsey’s own page did not return readable content to us, so they are cited as reported by Construction Dive rather than read at source. Various outlets have carried a dollar estimate attributed to the same report; we could not verify it against any source we could read, so it does not appear here. DEWALT’s survey did not disclose a sample size.
