Chris M. / 7 August 2026 / Reviewed 7 August 2026

Where AI Saves Time in UK Construction, and Where It Does Not

Take an illustrative case. An operations director is asked to justify a line in next year's budget. She searches for what AI does for construction businesses and finds, near the top of the results, the claim that construction workers lose around 35 per cent of their time to non-productive activity ⁸. It is a good number. It is on dozens of sites. It is also, when you follow it back, a self-reported survey of 599 people published by a software vendor in 2018 ⁸, and it will not survive 10 minutes with a finance director who decides to check.

Artificial intelligence saves time in UK construction on a narrow and identifiable class of work: converting unstructured material the business already produces, messages, photographs, recordings, scanned paperwork, into structured records somebody would otherwise have typed. It does not reliably save time on judgement, on reading drawings, or on any task where the answer has to be right first time without being checked. No trustworthy UK measurement exists of how much time construction actually loses to administration, and this guide does not invent one.

That last point deserves stating plainly, because the absence is the most useful thing on this page. Searches were made across the Office for National Statistics (ONS), gov.uk, the Health and Safety Executive, the Chartered Institute of Building (CIOB), the Royal Institution of Chartered Surveyors (RICS), the Federation of Master Builders and parliamentary publications. No UK statistical authority or professional body publishes a figure for the proportion of time construction workers or managers spend on administrative work. Every quantified claim traced back to a vendor's own survey.

The number you have probably already seen

The 35 per cent figure, and its companion claim of roughly 14 hours a week, originate in a 2018 report published by PlanGrid, a construction software vendor, in partnership with FMI Corporation ⁸. It is worth knowing exactly what it is before repeating it.

The study surveyed 599 construction leaders, of whom 500 were in the United States and 99 were spread across Australia, New Zealand, the United Kingdom and Canada combined ⁸. There is no separable UK sample. Each respondent estimated their own time across six activities, and the authors then grouped those activities into optimal, acceptable and non-optimal categories ⁸. The widely quoted headline cost figure is a modelled extrapolation whose formula the report states in its own footnote ⁸.

Three things follow. It is an estimate rather than a measurement, because people were asked to recall how they spend their weeks. The non-optimal grouping is the authors' own categorisation and includes conflict resolution and rework, which are not administration by any ordinary reading. And it is now eight years old, describing a period before the current generation of tools existed.

None of that makes the underlying intuition wrong. Anyone who has run a project knows the evenings spent writing up. It makes the number unusable as evidence, which is a different problem, and one that matters if the number is going into a board paper.

What the official statistics actually say

There is real UK data here. It is less convenient than the vendor survey and considerably more defensible.

The most recent official measure of construction labour productivity, ONS series DJP8, records output per hour falling by 1.8 per cent in 2024 and by a further 2.2 per cent in 2025 ². Two consecutive annual falls is the strongest single fact available on this subject, and it is a real problem statement rather than a marketing one. The honest caveat belongs immediately beside it: the series carries no causal explanation, and a fall in output per hour is not evidence that administration caused it ².

The longer run picture is consistent. ONS has stated that average productivity levels in the construction industry have remained consistently below the UK average and have grown more slowly, with civil engineering the most productive sub-industry and specialised construction activities the least ³. That analysis dates from 2021 and uses earlier data, so it describes a long-standing pattern rather than current conditions.

Scale and structure matter for why the problem is hard. Construction new work in Great Britain was worth £140,684 million in 2024. Separately, 370,770 registered construction firms were operating in Great Britain in the third quarter of that year, and construction related employees totalled 1.4 million, a figure that excludes self-employment and therefore understates the workforce ¹. Of the UK's 2.1 million companies, 44.0 per cent are single employee limited companies, and construction together with professional, scientific and technical activities accounts for almost two fifths of those ⁴. An industry built from very small firms keeps its information in individual heads, phones and inboxes, because there is nobody whose job it is to keep it anywhere else.

Where the documentation load genuinely comes from

Rather than guess at hours, it is more useful to name the work, because a reader can then count their own.

Some of it is a legal duty. For higher-risk buildings, duty holders must keep a digital record of building information, kept digitally, secure from unauthorised access, available when someone needs it, presented usably, and serving as the building's single source of truth ⁶. That applies only to higher-risk buildings and their named duty holders, and it creates no general paperwork duty ⁶, but where it bites it is not optional and it does not scale down.

Some of it is commercial pressure that shows up in surveys of the profession. In the RICS UK Construction Monitor for the first quarter of 2026, a sentiment survey of 938 chartered surveyors, planning and regulation was the second most widely cited obstacle to activity, referenced by 63 per cent of respondents, behind financial constraints at 66 per cent ⁵. Read that carefully, because it is easy to over claim: the respondents are chartered surveyors rather than contractors, the question is about obstacles to building activity rather than time spent on paperwork, and RICS states in its methodology that net balance data is opinion based and does not quantify actual change ⁵.

And some of it is the cost of things going wrong. The Get It Right Initiative estimates the direct costs of avoidable error in UK construction at around 5 per cent of project value, roughly £5 billion a year, with wider estimates ranging from 10 to 25 per cent of project cost ⁷. That is an industry initiative's informed estimate drawn from a study group of 17 organisations, based on research now roughly a decade old, and error is not the same thing as administration ⁷. It is offered as an order of magnitude for what poor information costs, not as a measurement of anything.

Where AI reliably returns time today

The work AI does well has a shape. It takes material that already exists in a form nobody can query, and turns it into something structured, and it does so at a quality that a human then checks rather than rewrites.

Transcription is the clearest case. A supervisor speaking into a phone for a minute or two produces a written record that would otherwise have been typed up in the evening or, more often, never typed at all. Extraction is the second: delivery tickets, invoices and certificates carry a small number of fields in predictable places, and pulling those fields into a register is the kind of task that pays back daily. Drafting is the third, where a monthly report assembled from the month's records gives somebody a first version to correct instead of a blank page.

What connects all three is that a human remains the last step. The vendors say so themselves. Anthropic's documentation states that its models might hallucinate or make mistakes when interpreting low quality, rotated or very small images, that object counts might not be precisely accurate, and that users should always carefully review and verify image interpretations and should not use them for tasks requiring perfect precision without human review ¹⁰. Google's guidance for its Gemini models tells users to rotate pages to the correct orientation before uploading and to avoid blurry pages ¹², which is the most practical sentence in this entire guide for anyone whose input is a photograph of a delivery ticket taken in a site cabin.

The ceilings are worth knowing before anyone promises a system that reads everything. Anthropic documents a maximum request size of 32 MB and a maximum of 600 pages per request, notes that the page figure is a ceiling rather than a promise because dense documents can fail earlier, and states that formats such as XLSX and DOCX must be converted before use ¹¹. Google documents its own different limits, 50 MB or 1,000 pages, and warns that anything other than a PDF loses its charts, diagrams and formatting and is treated as plain text ¹². Those numbers differ by vendor and must not be blended into one figure.

Where it does not save time, and one honest gap

Three categories resist, and being straight about them is what makes the rest credible.

The first is judgement. Deciding whether a delay is compensable, whether a design is buildable, or whether a subcontractor's application is fair, is not slow because of typing. It is slow because it is difficult, and a faster first draft does not make a hard decision easier.

The second is anything needing a guaranteed answer. Where an error has to be caught before it leaves the building, a system that is usually right imposes a checking burden that can consume the time it saved. A process that saves 20 minutes and requires five minutes of review saves 15, and an ROI case that omits the review time is describing an unreviewed system.

The third is drawings, and here the position is an absence of evidence rather than a vendor admission. No vendor documentation was found during this research making a verifiable capability or accuracy claim for extraction, classification or quantity takeoff from construction drawings. Microsoft's own layout model documentation states that embedded images are not supported, and that bounding regions for figures cover only the core content and exclude the associated caption and footnotes ⁹, which is a long way from interpreting what a drawing depicts. Products in this space are sold and some may work well; the guide to processes ranked by evidence sets out what the vendors do and do not document, in detail.

How to get a number you can actually defend

Since no national figure exists, the only credible number is your own, and producing one is not difficult.

Take two ordinary weeks and have the people doing a specific task note the time it takes. Not the team, and not in general: the diary took this supervisor this long on these days. Then count the timestamps you already hold, enquiry received against first reply, completion date against invoice date, over the last 20 instances. That is a baseline, and it costs a fortnight of light effort while the work is being scoped.

Do it before anything is built, because afterwards the before is gone and the only available answer is a reconstruction that deserves the scepticism it will get. Measure the task and never the person, or the notes turn into fiction within days. The fuller method sits in a separate piece on measuring automation in time rather than money, and the reason that piece argues for hours over pounds is exactly the reason this one refuses to quote a national percentage: hours are observable and contestable numbers are not.

Once you have hours, the arithmetic that turns them into a business case is a separate discipline with its own traps, set out in the guide to calculating return on construction AI. Which tasks are worth measuring in the first place is answered in construction processes you can automate today, and the question of what may safely go into these tools is covered in uploading construction documents to AI.

The limits of this guide

Three, stated plainly.

There is no UK measurement of administrative time in construction, so the central quantity a reader might have come here for does not appear. What is offered instead is the official productivity position, a description of where documentation load comes from, and a method for measuring your own.

The vendor capabilities described here were read on 7 August 2026 and this area changes quickly, so treat the specific ceilings as needing a check rather than as standing facts. The ONS productivity series has a further release due on 18 August 2026 ².

And nothing here establishes that AI improves productivity in construction at industry level. Two consecutive annual falls in output per hour ² is a description of the problem, not evidence about the solution. The case for the individual tasks above rests on their mechanism and on vendor documented capability, and any firm wanting evidence about outcomes will have to generate it themselves, which is the honest state of this subject in 2026.

Extraction and drafting of the kind described above is what our document automation work consists of, and it always starts by measuring the task rather than by choosing a tool.