What the thirty-nine per cent automation figure does and does not mean

‘Thirty-nine per cent of non-physical construction work is within theoretical reach of automation’ is a sentence the industry can now say without looking anything up. It turns up on slide four of a strategy pack. A board member reads it back slowly, because it is the only figure in forty pages anybody thought worth writing down, and a quarter later it is the premise underneath a budget line. Conference slides, vendor decks, board papers. It almost never arrives with a denominator, a year, a geography or a method section.

Treat it as a circulated figure rather than a census. The question worth asking is not whether thirty-nine is high or low, but what would have to be true of the labour, and of the issued set, for that ceiling to become production.

The claim under test is the whole sentence, not the integer. What would falsify the sceptical reading is a published method that defines non-physical work, defines theoretical reach, defines automation, names the population of hours, and then shows those hours leaving the job as issued work. Until that method is in the room, you are quoting a caption.

A ceiling drawn on somebody’s taxonomy

Most AECO numbers arrive without a denominator. Documentation volume without incident rates. ‘Thirty per cent of lifecycle data lost’ without the moment of loss. Automation percentages without a definition of non-physical work. Read the method. If there is no method, you are reading a brochure.

This one works as a slogan because it is large, odd and portable. Thirty-nine is more precise than thirty and less theatrical than fifty, and an odd integer travels as though somebody counted. Precision is not a method. The slide never has to say what was counted.

The figure sometimes travels with a workflow list attached — a hundred-and-fifty-workflow taxonomy somebody is said to have scored, most visibly in McKinsey's July 2026 AEC report. Treat that list exactly as you treat the percentage. If the room can produce the definitions, the scoring rule and the population of hours, you have a citation you can work with. Until then it is a gesture: a ceiling drawn on a consultant’s taxonomy and recited as though it forecast next Tuesday.

A ceiling is not a run-rate. A taxonomy is not a job. Theoretical reach is not an issued revision. Three limbs, glued together by one sentence, and each of them worth pulling apart.

‘Non-physical’ is whatever was left over

Non-physical is not a trade, a grade or a cost code. It is a remainder — whatever, in someone’s task list, is not pouring, fixing, lifting, isolating, setting out, driving or standing about in weather. The remainder gets treated as a single pile. Estimating, planning, document control, quantity surveying, design coordination, commercial correspondence, inspection paperwork, the minutes, the register, the claim file: all of it becomes the numerator’s home.

That pile is not one kind of labour. The estimator pricing a package and the document controller issuing a revision do not share a medium simply because neither of them is tying steel. One produces a number that may never become an instruction. The other produces the instruction the site is allowed to build from. Collapsing the two into ‘non-physical work’ is how a ceiling acquires a denominator big enough to survive.

The cut also buries the split you can see in labour-exposure research. Construction as an occupation sits low on current model coverage because most of the hours are presence. Construction as a firm contains office hours that look a great deal like knowledge work. A circulated automation percentage that starts counting only after the physical hours have been removed is not contradicting that ranking; it has changed the population. The site has left the frame; the trailer and the regional office remain. That is a different claim and ought to be stated as one.

Ask who was left in the remainder. If the method cannot say, non-physical is a caption for the hours everybody already suspected a language model could touch.

‘Within theoretical reach’ describes overlap and nothing else

Theoretical reach is a statement about overlap. Not adoption, and certainly not issue. A task can overlap a model — draft it, summarise it, classify it, retrieve it, reconcile a register against a pile — and still never become the way issued work actually gets done. Reach is a matching exercise. Production is a dated act: named containers, a revision, a suitability or purpose, a list of recipients, and a predecessor that has been withdrawn.

The ISO 19650 states are unromantic about this. Work in progress is not Shared. Shared is not Published. A chat window pointed at last Thursday’s export is reading a picnic of files rather than the set the hut is allowed to plot, and theoretical reach can be entirely true of the picnic while remaining false of the issued set.

What usually happens next is that the ceiling gets read as a pipeline. Thirty-nine per cent is available; available becomes this quarter; this quarter becomes a pilot. The pilot that dies is what reach looks like when somebody treats it as a production plan — no pin to the issued set, nobody owning the exception, Tuesday still running around the tool. The funeral costs the figure nothing. It reappears in the next steering pack as evidence that the opportunity is still large.

A taxonomy can score a workflow as reachable because a model can draft the artefact. The job can still issue that artefact off a USB stick, out of an inbox, or from a plot that was superseded on Monday. Reach without issue is a demo. Issue without reach is how most buildings still go up, and the slogan only has a number for the first of those.

What counts as production on a given job is boring and local. Name the workflow. Name the issued containers it is allowed to see. Name the person who can stop a wrong output. Count whether Tuesday still runs around the tool. If any of those answers is a slide, you have not left the taxonomy.

‘Automation’ is doing an enormous amount of lifting

In the circulated sentence, automation usually means no more than that a model can do part of the task. It does not mean the labour has left. The unofficial set in the trailer has not been withdrawn. The document controller has no fewer supersessions to chase on Friday afternoon. The claim file still contains no contemporaneous record. A draft nobody will sign is one more file in the picnic.

Issue is the act the site already understands: these sheets, this model state, this revision, released to these companies, for this purpose. A tool that cannot answer those four questions is retrieving over a folder. Retrieval earns its keep often enough, but calling it automation lets a summariser borrow the glamour of a robot.

There is a second confusion buried in the word. Partial assistance gets scored as though it were substitution. A planner who pastes last week’s minutes into a window has not automated planning. A quantity surveyor who asks a model to list the RFIs in a ZIP has not automated measurement. The hours around the prompt — checking the revision, arguing the clause, walking the slab, refusing the wrong depth — are still the job. If the taxonomy scored a task as reachable because the artefact happens to be a file, it scored the medium rather than the labour. The file is what remains after the labour, and any ceiling that stops at the file will always look larger than the hours a competent person still has to spend before anyone is allowed to build.

Limb of the claim	What it actually says	What it is often read as
Non-physical	A remainder — whatever isn't pouring, fixing, lifting, isolating, setting out, driving or standing about in weather	A defined trade or cost code
Theoretical reach	Overlap — a task a model could draft, summarise, retrieve or classify	Adoption already achieved
Automation	A model can complete part of the task	The labour has left, or the job is done

What the number does not prove

It does not prove that thirty-nine is the right number on your job. It does not prove that a named house counted a hundred-and-fifty workflows, or that anybody in your building has seen the list. It does not prove a headcount reduction, a programme saving, or a date by which the trailer will finally trust the official cupboard. It does not prove that software, a mandate or a model will save the outcome by itself.

What it proves is that the claim needs a source, a limit and a next measurement.

The measurement is unglamorous, which is presumably why it never reaches a slide. Define the population: the non-physical hours that already touch an issued container — transmittal, RFI, inspection, daily report, published model state. Define automated: the issued path now runs through the tool, an owner is named for exceptions, and the old path has been withdrawn from the people still quietly using it. Count those hours at the start of a quarter and again at the end. The gap is your number. It will not be thirty-nine, except by accident, and it will point at a pin, an owner, or a workflow that never left the slide.

Falsifying the slogan as a forecast is easy to state and rare to see. Produce, for a named population of hours, a method that moved those hours from theoretical overlap into issued production, and keep the unofficial set from growing in the gap while you do it. One good job does not retire the circulated figure. It does show that the figure was always a claim about a taxonomy rather than a constant of construction.

So when the sentence comes round again, do not argue with the integer. Ask three things and have them minuted. Which hours are in the denominator — is the site inside this population at all? How many of those hours currently pass through an issued container rather than an inbox? And what exactly are we proposing to withdraw, given that nothing is automated while the old path is still open to anyone who prefers it?

The third question is where the money is, so be straight with whoever holds the budget. Moving a workflow from theoretical reach to issued production is mostly not a licence cost. It is making the CDE states honest, making the transmittal record trustworthy, finding somebody senior enough to withdraw the unofficial set, and spending a quarter on tedium that will photograph appallingly. Clients and appointing parties can force a good deal of that through the information requirements they write into appointments; a contractor can only impose it on itself. In the meantime, quote the sentence whole — thirty-nine per cent of non-physical construction work within theoretical reach of automation, a circulated ceiling rather than a production forecast. The qualifiers are where the argument lives, and a board that hears all of them asks far better questions than one that only hears the integer.