A widely circulated construction AI statistic comes from Australian and North American fieldwork, and nobody checking it noticed.
On 17 August, a German trade publication for craft businesses ran a number that has since travelled well: workers in construction lose around 115 hours a week searching for data. It is a good number. It is specific, it is uncomfortable, and it converts instantly into a business case for whoever is selling document management software that morning. It comes from a Deloitte study published on 3 July 2026, commissioned by Autodesk. The study is real, the methodology is unremarkable in the good sense, and the fieldwork was done in Australia.
That last part is the whole story. It is also the part that never makes the trip.
Five bad numbers, four different failure modes
Over the past months we have had to discard four other widely circulated figures about German industry, each for a different reason. One rested on a sample of roughly thirty respondents and got quoted alongside a KfW survey with several thousand. One was attributed to a software vendor when it actually came from a Singaporean training provider. One described a self-selecting cohort of early adopters and got read as a population average. One carried a government sender that implied a scope the underlying programme did not have.
Each of those had a tell. Small sample, wrong author, biased cohort, misleading letterhead: all four are things a careful reader catches by looking harder at the source. The Deloitte number has none of them. The sample is large. The author is named and reputable. The client relationship is disclosed. The publication date is six weeks old. Every quality signal a diligent person checks comes back green.
It fails on one axis only, and it is the axis nobody checks, because geography is not usually where a study goes wrong. A number about construction feels like a number about construction. Concrete cures the same way in Brisbane and in Bochum. The temptation to treat the finding as portable is not stupidity, it is a reasonable assumption applied to a case where it does not hold.
Why the assumption breaks in construction specifically
Data search time is not a property of building. It is a property of the paperwork around building, and paperwork is national. Australian construction runs on different procurement law, different contract standards, a different digital baseline in the supply chain, and a different distribution of firm sizes. Deloitte's own work for Autodesk reports 46 percent of surveyed firms already using AI in some form. Against the German picture, where KfW's population-level surveying of the Mittelstand has consistently shown adoption in the low double digits and a majority of firms without any digitalisation plan at all, that 46 percent is not a benchmark. It is a contrast.
Used as a contrast, it is genuinely useful. It says: here is a construction sector on the other side of an adoption curve, and here is what changes when you get there. Used as a benchmark, it produces a specific and expensive error. A managing director in Lower Saxony reads that his peers are at 46 percent, concludes he is behind, and buys the tool the article was implicitly selling. He is not behind in the way the number suggests. He is in a different market with different constraints, and the honest version of his situation is harder and more interesting than the imported one.
The 115 hours figure has the same problem in a sharper form. Search time depends on where documents live, which depends on what the supply chain has already digitised, which depends on who the general contractors are and what they mandate. Transplanted to a German subcontractor with fifteen employees and a fax machine that still gets used, the number is not too high or too low. It is simply about somebody else.
The rule that has to survive contact with a deadline
None of this makes Deloitte's work bad. It makes it foreign, and foreign is a category most citation practice has no slot for. Publications check whether a study exists, who paid for it, and how many people it surveyed. Almost nobody checks where the respondents were standing.
The practical move is to treat the country of fieldwork as a mandatory field, exactly like sample size and sponsor, and to refuse to quote any figure whose geography you cannot state in the same sentence you quote it in.
Written out, that discipline is cheap. The 46 percent stays usable, as long as it always appears as "46 percent of Australian construction firms surveyed by Deloitte for Autodesk". The 115 hours becomes almost unusable in a German context, not because it is wrong, but because nobody has measured the German equivalent and pretending otherwise is a fabrication with a footnote attached.
What is actually missing
The uncomfortable part is what the discipline reveals. Strip out every foreign figure from the German construction AI conversation and what remains is thin. There is KfW on digitalisation, there are labour market numbers on the roughly half a million missing workers, and there is very little measurement of what AI adoption actually looks like on a German site. The imported statistics are not filling a gap in rhetoric. They are filling a gap in evidence, and they are doing it badly enough that the gap stays invisible.
Someone will eventually run that study. Until then, the most valuable thing anyone in this market can say about German construction AI is a specific number from a single named firm that they watched with their own eyes.