What moved
July is the month a percentage starts travelling as though it were a programme. The shape of this one is consistent enough to describe without pretending to a title page: thirty-nine per cent of non-physical construction work is said to sit within theoretical reach of automation, across a taxonomy of about 150 workflows. The sentence that leaves the briefing is the thirty-nine. The sentence that should leave it is everything after the comma.
There are three layers to that figure, and they get flattened in roughly that order.
The first is the percentage. It is a filing category. It is not a count of hours already automated, seats already sold or containers already published, though it is large enough to look like a sector event and precise enough to sound as though somebody measured it. A month in which the industry recites 39 per cent is not, by itself, a month in which anyone's issued set got shorter. The number is a ceiling on someone else's taxonomy, and a ceiling is not a forecast of next quarter's labour.
The second layer is the cut. The figure is not about construction as an occupation. It is about the non-physical slice — estimating, planning, coordination, document control, quantity work, reporting — the hours that already look like office labour inside a firm whose average hours do not. That matters, because the occupational-exposure research putting construction near the bottom of current AI coverage is still in the room from the spring, and people are trying to hold both sentences at once. They are not a contradiction. One averages the site and the trailer together; the other carves the trailer out and scores only the file-work. Reciting the thirty-nine as though it covered the hoist is a category error, and a familiar one at that — a headline bucket with no method behind the numerator.
The third layer is the taxonomy, which is the object that would make the percentage honest and the object almost nobody quotes. One hundred and fifty workflows. Theoretical reach is a matching exercise: a task is listed, a current or assumed model capability is listed beside it, and the output is the share of the list that looks, on paper, automatable. That output is not a share of a live job already automated. It has not been contracted. It has not been issued. A workflow that never touches an issued revision can sit comfortably inside the thirty-nine while the plotter down the corridor still takes a USB.
What this looks like on a job is a great deal quieter than the slide. A document controller is asked whether 39 per cent of her week can now be a model. An estimator is asked whether a taxonomy item called take-off is the same object as last Thursday's bill. An appointing party is asked whether the exchange information requirements ought to say "automated" because a mid-year briefing used the word. None of those questions is answered by the percentage. They are answered by which of the 150 items — assuming anybody has read them — change a published container, a suitability code, or the labour of knowing the current revision.
A pitch that says the industry can now automate two-fifths of the office is weather. What moved in July is a circulated ceiling, a non-physical cut, and a taxonomy that almost nobody opens.
What changed in the standards
A percentage is not a publication, and nothing in the series moved because one circulated.
No new part of ISO 19650 was published in July. The published series remains what it was: Parts 1 and 2 from 2018, Part 3 from 2020, Part 4 from 2022, Part 5 from 2020, and Part 6 on health and safety information from January 2025. Appointments that say "ISO 19650 aligned" are still aligning, if they align to anything, to those texts. They were not new this month. They were in use this month, which is the fact that governs your week.
Those texts already assume an appointing party, appointed parties, an information protocol, a common data environment and a published container. None of them assumes that 39 per cent of non-physical work has been automated. Nor, for the avoidance of doubt, does any of them forbid a model from drafting a transmittal; what they require is that information requirements, exchange and security be named. A chat window does not issue. Work in progress is not Shared. Shared is not Published. Theoretical reach moves none of those states.
The revision conversation around Parts 1, 2 and 3 is weather rather than a published replacement. Working groups collect comments; national mirror committees circulate drafts internally. Where the drafting had actually reached by the end of July, I cannot tell you with a clause number, and I would treat with caution anybody in a webinar who says they can.
The Building Safety Act's information duties in England were already in force. They did not arrive because a slide mentioned a golden thread, and they still want an identifiable, available, change-controlled record — not a prompt history that summarised last week's folder, and not a taxonomy item that never named a revision. Gulf permit-stage BIM requirements remain a class of public-record rule where the text is public; I have no July circular to cite and will not manufacture one.
What July does produce, a fortnight into the recitation, is a crop of unofficial text. A BIM execution plan with a new "automation" paragraph lifted from a briefing. An exchange information requirement that says "AI-enabled" because 39 per cent sounded like a method. A tender reply treating a chatbot pointed at last week's export as though it were a revised part. None of those is a clause, and all of them will be quoted back at somebody in a meeting.
A pitch calling a platform automation-ready because a circulated figure used the word theoretical is not a clause either. Openness that a party outside the authoring licence cannot exercise is still a brochure word, whatever a note called the month. Percentage season is not a standards event.
One thing to watch
Watch who opened the 150. That is the whole month in one question. If your appointing party issued a one-page note saying that this appointment does not rewrite its information requirements because a circulated figure used the word automation, somebody did useful work in July. If the supply chain added "automated workflows" to the execution plan and nobody can name which workflows, let alone which of them touch an issued revision, the job has bought a rumour. And if three applications now claim to cover the thirty-nine while the plotter still takes a USB, the record did not move.
Watch the phrase theoretical reach and keep it in its box. Theoretical reach is not deployed reach; deployed reach is not contracted reach; contracted reach is not a published state. If the figure stays a ceiling — a list of tasks that look, on paper, as if a model could touch them — fine. If it becomes a programme, a saving or a headcount line before anyone has mapped the list onto this job's containers, someone has bought a caption.
Watch the non-physical cut the same way. Trailer hours and site hours are not one occupation for the purposes of a model. A figure that scores only the file-work does not retire the hoist, and it does not retire the unofficial PDF set either. There is an irony worth sitting with: the hours that look most automatable on a taxonomy are very often the hours construction already fails at without any model at all — finding the current revision, reconciling a register, proving who was issued what and when. A model pointed at last Thursday's export is reading the exposed slice and missing the issued set.
Two weeks after the briefings, the test is not whether anyone can still recite the number. Everyone can. It is whether the argument in the document control office has changed shape.
Here is a cheap experiment for August, and it costs an afternoon rather than a budget. Take five of your own workflows — the five that eat the most hours, not the five that sound most modern — and write down, for each, which published container it touches and at what suitability. Then ask anyone selling you automation to point at where in that list their product sits. If they can, you have something to price. If they can only point at the thirty-nine, you have learned that the taxonomy was never opened at their end either.