An independent magazine about AI at work in construction. Every story starts on a real job site, with the people who used the technology and the results they will answer for.
McKinsey Puts 39 Percent of Construction Desk Work Within Reach of AI | ConstructionMagazine.ai
McKinsey Puts 39 Percent of Construction Desk Work Within Reach of AI
McKinsey's estimate and DEWALT's survey measure different things. The contractors who have published their own numbers count weekly use, in hours, on a single tool, and they are the large ones.
McKinsey published its paper on AI in architecture, engineering, and construction on 15 July 2026. The number that traveled was 39 percent: the share of construction's nonphysical work the firm judges automatable, against 50 percent for architecture and engineering. McKinsey reached it by mapping 150 workflows across 25 domains. It puts bid-or-no-bid analysis, estimating, and proposal drafting in the first 18 months, work built on proprietary data such as RFIs, drawings, and specifications in the next four years, and autonomous equipment after that.
A second number arrived nine days earlier and has traveled with the first ever since. Nine percent of construction professionals told DEWALT they use AI in their day-to-day work, according to Daily Commercial News. Another 39 percent were piloting tools and 35 percent were researching them. The survey covered more than 3,400 professionals in six countries. The same fieldwork has appeared as 8 percent in DEWALT's own April release and as 16 percent in a May report, so the nine carries a margin that neither headline shows.
The numbers do not conflict. One is a forecast of what could be automated and one is a count of who opened a tool. Other surveys fill in the middle. BuiltWorlds asked 53 people at 30 organizations, most of them large general contractors: nearly two-thirds rated their AI maturity as average, and the blockers they named were data privacy at 62 percent, missing internal skill at 58, and poor data at 56. Dodge's survey of more than 230 contractors found 86 percent of large firms expect AI to be a competitive advantage against 69 percent of smaller ones, and cost was a concern for 49 percent of the smaller group and 26 percent of the large.
The forecast and the count describe different things. AI-assisted illustration.
How each number was counted
The four figures come from four different instruments, and the instrument decides what the number can say. McKinsey's 39 percent is not a survey at all. Analysts listed 150 workflows across 25 domains of construction work, judged which steps a model could perform, and weighted the result by the time those steps take. It measures the shape of the work, not anyone's behaviour. DEWALT's nine percent is a questionnaire answered by more than 3,400 people in six countries, and the word that matters in it is daily. The same respondents produced a higher figure when the question was about any use, which is why the fieldwork has surfaced as 8, 9, and 16 percent in three releases.
BuiltWorlds asked 53 people, most at large general contractors, so its two-thirds average-maturity figure describes the firms that already have technology staff to answer the call. Dodge's 230 contractors were asked about expectations, not use. Put side by side, the numbers do not disagree so much as describe different rooms: a consultant's map of the work, a tradesperson's morning, a technology director's self-assessment, and an executive's forecast. A reader who wants one number should first decide which room they are in.
What a measured number looks like inside one firm
Survey percentages describe an industry. The contractors that have measured their own use describe something else. One general contractor of about 700 people reports that 72 percent of its staff used AI in a given week, up sharply over a few months, and it knows this because the number comes from telemetry rather than a questionnaire: roughly a thousand hours of assisted work in four weeks, counted on one licensed tool and excluding the others in use. Weekly is not daily, one tool is not all tools, and a self-reported survey answer depends entirely on which tool the respondent had in mind. The nine percent and the 72 percent are not measuring the same thing, and neither is wrong.
The same firm's longest-running use is the one McKinsey puts first. Its preconstruction group has used AI-assisted takeoff for a couple of years and estimates the saved time at a few hundred thousand dollars, with most of the clicks gone from the process. Its technology lead is also candid about a category the surveys count as adoption: a cost-history application he built for the firm in an afternoon that can produce a conceptual estimate and benchmark it by region, and that he does not count as a product. Data in a tool is not the same as a tool the firm relies on.
Telemetry changes what a percentage means. A licence server counts sessions and hours whether or not anyone remembers to report them, so a weekly-use figure from telemetry is a floor for that tool and a ceiling for nothing else. It cannot see the free chatbot on a project engineer's phone or the takeoff assistant that runs inside another product. The firm above counts one licensed product and says so. That candour is the useful part. A firm that publishes a usage figure without saying which tools it counted has published a number nobody can compare.
The question before the tool
Consultants who take calls from mid-size contractors describe why daily use stalls at a chatbot for correspondence. Their first question to a firm asking for AI is not about models. It is whether the firm's data is any good: whether every project manager fills in the submittal log the same way, whether the cost codes on one job mean what they mean on the next. A company dashboard built on inconsistent inputs is worthless, and a model pointed at it is worse than worthless, because it answers fluently.
The examples are mundane and expensive. A contractor that grew from about $40 million to $150 million in three years runs five systems for accounting, project management, purchasing, service, and time. A purchase order created in the purchasing tool adds sales tax and pushes a commitment with the tax into the project system, but the integration into accounting carries the line items and drops the tax. The fix is custom middleware, not a model. At another firm, the connector between Procore and the accounting system has to be restarted every morning at 9:30 or it stops. Consultants who work with these firms put a threshold on where AI adoption is even plausible: above $50 million in revenue, firms are buying and integrating systems. Below about $15 million, they rarely have anyone whose job it is to think about technology at all.
The connector problem has a daily rhythm that anyone who has run a mid-size contractor's back office will recognise. A link between the project system and the accounting system that has to be restarted every morning before it fails is not an edge case. Consultants describe it as the common state of integrations bought one at a time over five years by five different managers. A model asked to summarise the day's costs from a record with a dropped sales-tax line and a connector that stopped at 9:30 will summarise them confidently. Nine percent daily use may be the rational number for an industry whose data pipes are in that state.
The readiness check consultants run before an AI purchase
1Name the one report or output you need
→
2Find which system each input lives in
→
3Check that every project collects it the same way
→
4Bring the sources into one enterprise data layer
→
5Point the model at that layer
→
6A person verifies the number before it leaves the company
Adoption is not productivity
At the top of the market the numbers invert. A contractor with about $5 billion in revenue gives each project manager a technology package of roughly forty tools, and each superintendent a few less, and its chief executive says the firm has not seen a productivity benefit from the count. He also describes internal adoption of a general-purpose model as rocketing, without a figure. A large contractor can report high adoption and no productivity change at the same time. That is not a contradiction. It is what adoption looks like before workflows change.
The denominator is the part the surveys skip. Most construction firms have fewer than fifty people. They do not buy enterprise software, they do not integrate five systems, and no consultant is calling them about their data. Any industry-wide usage percentage is dominated by them, and the contractors who publish measured figures are not among them.
Two large contractors have published their own numbers, as relayed by the analyst Gerard de Valence. Turner Construction said its staff built more than 400 custom AI applications under an OpenAI partnership and claimed more than 70,000 hours of annual productivity gains. Balfour Beatty spent £7.2 million on Microsoft Copilot for 13,000 UK staff and reported that more than half had used it at least once in the first 12 weeks. Both are self-reported. Both count something different from the nine percent.
The field tools with customers who will say so this year are cameras, capture, and a few retrofitted machines. The desk tools this magazine has seen in use check estimate arithmetic, compare drawing sets, and review subcontractor proposals. None of them shows up as a percentage, and the percentages that exist describe an industry mostly made of firms nobody has surveyed.
How to read the next adoption number
Ask what use means in the question: daily, weekly, ever, or piloting. The same fieldwork can produce 8, 9, or 16 percent.
Ask who was surveyed and how many. Fifty-three technology leads at large GCs and 3,400 tradespeople describe different industries.
Ask who paid for the survey and what they sell. A tool vendor's adoption figure is a marketing number until someone else reproduces it.
Prefer a firm's own telemetry, and ask which tools it counted, because one licensed product is never all of them.
Separate adoption from productivity. A firm can report high use and no measured gain in the same sentence, and several do.
The 39 percent will be quoted for a year as a promise and the nine percent as a rebuke. Both are honest measurements of different things. The only figure a contractor can act on is its own, counted on its own tools, against its own hours.