The Buildots Intelligence Lab announced its inaugural Global Construction Benchmarks Report on 8 October, datelined “TEL AVIV, Israel, October 8th, 2026”. The release says the benchmarks are “based on anonymized, aggregated data from hundreds of global construction projects across five verticals,” and gives the following figures for a typical week:
- 29.3% of construction activities start on time. The release states the complement directly too: “70.7% of construction activities miss their planned start date.”
- 52.8% of active construction work stays on schedule.
- 37% of planned weekly work goes undelivered — “The typical project delivers just 63% of its planned output.”
- The typical activity “finishes 5.3 weeks late,” against 2.3 weeks for top-quartile projects.
- Of the 21 activities measured, 20 fell short of plan, “ranging from 38% (resilient/soft flooring) to 122% (drywall framing, the only activity to exceed plan).”
Every one of those numbers is Buildots’. None can be independently verified from outside the company’s dataset, and this piece makes no attempt to: what follows is about where the numbers come from, which is the part that makes them worth writing about.
Measured, Not Surveyed
Construction’s productivity statistics are mostly self-reported. This newsroom has covered several of them this year, and the sourcing is consistent: ServiceTitan’s Commercial State of the Trades, which found that 62% of commercial contractors have tried AI while 15% of users could show clear ROI, rests on a July survey of 1,020 contractors. McKinsey’s finding that AI could automate 39% of construction’s non-physical work is a study estimate set against survey data showing single-digit daily use. Suffolk and MIT’s estimate that six AI levers together could take 17–20% off cost came, as we wrote at the time, from a model, not a jobsite.
The Lab’s report landing page makes exactly that contrast its pitch: “Many widely cited construction statistics rely on self-reported surveys or retrospective reporting. This report is built differently.” It says the data is “filtered against strict data quality thresholds,” that work is measured “activity by activity, element by element, week by week,” and that the production metrics extend to “the feet, square feet, and pieces actually installed each week across 21 core construction activities, benchmarked against the rates their schedules assumed.”
That is a different kind of claim from a survey finding, and it is the reason the report matters even to a reader who discounts every headline percentage in it. Buildots’ platform works by comparing site capture against the model and the programme; the by-product is a weekly record of planned-versus-actual at activity level across a portfolio of projects. Publishing from that by-product is a choice, and it is the choice that distinguishes this release from a vendor survey.
Amir Berman, Buildots’ VP of Industry Transformation and head of the Intelligence Lab, frames the gap in the release this way: “Everything else that goes into a building has long been standardized – equipment, materials, specs. But one thing that hasn’t been is the work itself.” Roy Danon, the company’s co-founder and CEO, adds: “You can’t fix what you can’t measure, and the construction industry has always lacked a source of macro-level truth about its own performance.”
Both statements are the company’s characterisation of the industry, not findings in the data. Rate books have given estimators theoretical crew production rates for decades — the release acknowledges as much, and its argument is about continuous measurement against the plan rather than about the absence of production rates altogether.
The Interesting Number Is the Spread, Not the Average
A typical activity finishing 5.3 weeks late is a bad headline for the industry. The 2.3 weeks that the release attributes to top-quartile projects is the more useful figure, because it is the same trades doing the same work under the same conditions. Danon makes that point in the release: “This report proves strong performance already exists in the industry, it’s just inconsistent.”
The same shape appears in the production table. Twenty of 21 activities under-delivering against plan could be read as an industry failing to execute, or as an industry whose schedules are systematically optimistic — the release does not adjudicate, and neither will we, because the distinction cannot be settled from a benchmark table. What the 38%-to-122% range does establish is that the gap is not uniform: resilient and soft flooring delivering 38% of planned output and drywall framing delivering 122% are not the same problem, and an owner told that “construction runs late” learns less than a planner told which activities to buffer.
The release lists what the activity-level benchmarks cover: “schedule adherence, on-time starts, completion overrun, planned-vs.-actual production, peak-vs.-typical performance, and trade-level reliability across 21 core construction activities.” Peak-versus-typical output variance is the one worth watching — how much more a site builds in its best week than in an ordinary one is the kind of number that argues directly about whether a programme’s float is real.
Who Is Publishing, and Why It Matters
The Lab describes itself in the release’s boilerplate as “construction’s first AI-powered research hub” — Buildots’ superlative, reported here as the company’s claim. It says the Lab “operates as an autonomous research unit within Buildots” and “invites industry professionals, academics, and analysts to challenge it with their own questions and hypotheses.” The next editions are already named: “sector-specific and regional editions – starting with data centers and healthcare.”
That roadmap is commercially legible. Buildots raised $130 million on 14 September, a round whose announcement named Intel, Digital Realty, Turner, JE Dunn, HOCHTIEF and Bouygues among its users, and whose logic — we argued at the time — rested on whether the accumulated record of how projects actually progress is a defensible asset or a temporary one. We wrote then that the company’s framing was “a plausible argument” but “not yet a proven one.” A benchmarks series, with data-centre and healthcare editions to follow, is the company making that argument in public: the dataset is valuable enough to publish from, and publishing from it is how you get owners to treat your numbers as the reference.
It also means the reference being established is the vendor’s. There is nothing improper in that — the release is explicit about whose data it is, and the Lab’s stated invitation to be challenged is the right posture — but a benchmark becomes authoritative by being cited, and the company that owns the measurement instrument is the one publishing the scale. The release does not describe the sample’s relationship to the company beyond calling it “anonymized, aggregated data from hundreds of global construction projects,” but the data is the platform’s own output, so a reader pricing a programme against “52.8% schedule adherence” is pricing it against sites that were being captured this way.
What the Release Does Not Say
The release gives no project count beyond “hundreds,” and it does not say which five verticals the data covers. Sectors do appear elsewhere in the document — data centres and healthcare as the next editions, and a boilerplate line describing projects “from local schools and multi-family projects to the world’s most complex data centers and mega-projects” — but neither list is presented as the five. It does not state the period the data spans, describing it only as “recent industry performance data,” and it gives no geographic breakdown and no definition of how “planned output” is derived from a programme. Nor is there anything, in the release or on the landing page, putting numbers behind the landing page’s “strict data quality thresholds.” The full report — “including sector breakdowns, quartile data, and detailed production-rate tables for 21 core construction activities” — is published at the Lab’s landing page, but it is gated: the page carries no link to a file and instead loads a HubSpot form under the heading “Get the guide.” So the methodology section is not something this piece has read. Everything attributed above comes from the release and the public landing page, and is labelled as the company’s figures throughout.
