Chris M. / 13 August 2026 / Reviewed 13 August 2026

The Construction Automation Audit

This is not a list of things to automate. It is a method for finding the ones that matter in your business, which cannot be done from the outside.

The audit produces a ranked list of recurring tasks, each with an owner, a frequency, an hours-per-year figure, a count of how many people do the same thing, and a verdict on whether it needs judgement or is simply information being moved. It is done before any tool is chosen, because the answers determine what to buy, and because a business that has not measured where its hours go cannot tell whether a system worked.

RICS found that 29% of respondents measure productivity quarterly while 14% never measure it at all, and that no clear definition or globally agreed framework exists for improving it ³. An audit is how a business stops being in the second group.

What it will cost you to run it

Not an afternoon, if you do it properly.

Estimates collected from the person who requested a task are reliably wrong, so you have to ask the people who do it: project QSs, document controllers, site managers, bid coordinators. Across a couple of regions that is two to three weeks of elapsed time and a nominated owner, not a lunch break. Budget for that or the data will be thin.

There is a second problem, and it decides whether the numbers mean anything. Anyone who suspects an automation review is a headcount review will inflate their hours defensively or deflate them to look efficient. You cannot correct for that afterwards. You defuse it in advance, in writing, from a named person, before the first conversation.

Something like this, which a director can send on Monday without writing it:

We are mapping recurring admin across the business to work out what we should stop doing by hand. This is not a headcount exercise and no role is under review. I need an honest picture of where your week goes, including the parts that feel too small to mention. If a task takes you three hours and you think it should take twenty minutes, that is exactly what I want to hear about.

The six questions to ask of every recurring task

One row per task, in a spreadsheet. Resist evaluating anything while you are still collecting.

Rank by hours per year multiplied by the number of people doing it, then weight by seniority. Ranking on seniority alone is the common mistake and it points you at the wrong work: a director's task is expensive per hour but irreducibly one-off, while a document controller's task is cheap per hour and fixing it once fixes it thirty times.

A starting list, so nobody faces a blank page

The hardest part of any audit is the first column. Edit this rather than inventing it.

Valuations and applications. Payment notices and pay less notices against the contractual clock. Subcontractor order and payment administration. CVR assembly. RAMS review and chasing. ITP sign-off. RFI and technical query logs. Site diaries. Progress reports. Handover O and M manuals and the health and safety file. PQQ and standard questionnaire responses. Framework KPI returns. Plant and labour returns. Defects and snagging lists. Discharge of pre-commencement conditions. Section 38, 278 and 104 technical approvals. Retention release. Final account close-out. Board and monthly business review packs. Subcontractor prequalification. ISO audit evidence. Payment practices reporting. Tender quote levelling. Bid case studies and project data sheets. Meeting minutes and actions.

The judgement test, which is the one that matters

Most tasks are a mix, and the mix is the finding. Split the task rather than judging it whole.

Be careful here, because the tempting version is wrong. A project QS preparing a CVR is not doing retrieval and then thinking. Deciding what to accrue for a subcontractor who has not applied, what to carry for an instructed but unvalued variation under JCT or compensation event under NEC, and what the forecast final account looks like, is commercial judgement exercised at the moment of assembly. The genuinely mechanical parts are narrower: pulling committed cost and subcontract liabilities from the ledger, reconciling against the last period, and formatting.

So record two numbers per task:

• Assembly: retrieval, formatting, reconciliation, chasing, filing, renaming, rekeying.
• Judgement: interpretation, decision, challenge, the conversation that follows.

A director who says a board pack takes four hours is usually describing forty minutes of assembly, two hours of waiting on other people, and an hour of deciding how the difficult job gets worded. That middle number matters, because chasing is automatable and deciding is not.

The rule that comes out of this is short enough to hold in your head. Automate the work around the work. Do not automate the judgement, and be suspicious of anything sold on the basis that it will.

The question most audits never ask

If those hours came back, what would we want that person doing instead?

An audit that stops at hours saved produces a cost reduction. One that answers this produces capacity, and a construction board values those differently.

The evidence supports the caution. Fewer than one in six of the 3,500 UK businesses DSIT surveyed use AI at all, and even among that minority the productivity story runs ahead of the revenue one: three quarters report a workforce productivity gain, while over three quarters have seen no change in revenue ⁶. Both figures are self-reported perceptions rather than audited outcomes, and the gap between them is the thing to plan around. Recovered time does not convert itself into money.

Write the answer per task, in advance, and be specific. Not "more strategic work" but "two client visits a month", "walk every project over £2m once a fortnight", "review every tender above a threshold before it goes out".

Who pays for it

Decide this early, because in a contractor it determines whether the pilot happens at all.

If the cost lands in prelims, the project QS will resist it, because it hits their job's margin for a benefit shared across the business. If it lands in overhead, the finance director resists it as unallocated cost. Neither objection is unreasonable. The workable answer is usually overhead for the pilot and a decision at the end, but the point is to have the argument before the work rather than in month three.

What to do when IT says no

This is what actually stops these initiatives, not the business case.

Most project data sits under client confidentiality provisions, inside an ISO 27001 scope, and increasingly under Building Safety Act record-keeping duties. Before anything reaches a third-party model, your IT director and general counsel have a veto, and they are right to.

Have four answers ready. Where the data physically sits and under whose contract. Whether anything is used to train a public model, and how that is evidenced rather than asserted. Which client contracts contain confidentiality provisions that bite, checked rather than assumed. And who the named human reviewer is for every output, because "the system produced it" is not an answer to a client or an auditor.

Do that work first and the conversation is short. Skip it and the audit becomes a document nobody acts on. The same discipline applies to the common data environment: on a job where the client mandates the CDE, you are working inside their governance, not designing your own.

Scoring what you find

Filter the ranked list through three tests before choosing anything.

1. Is the output already defined? A task producing a document nobody has specified is not ready. Fix the specification first, or you will automate an argument.
2. Is the source data reachable? If the information lives in an inbox, a phone or somebody's head, the first job is getting it somewhere a system can see it.
3. Is there a named reviewer? Every automated output needs a person accountable for it. If nobody will own the review, the task stays manual.

The third is where most candidates fail, and it should. Global research from RICS puts AI use at zero in 45% of organisations and fully scaled in just 1%, and names cost and effort, a shortage of skilled people, and inconsistent approaches across the supply chain as the reasons why ⁷. Those are readiness problems, and an audit that surfaces them before a purchase has paid for itself.

What this audit will not tell you

Five things, better known before you start.

It will not tell you what a system costs. It ranks candidates; pricing comes after.

It will not tell you the fix is technology. UK SME data from the Enterprise Research Centre shows the payoff depends heavily on which tool you pick, not how many: pairing VR and AR with automation tends toward diminishing returns rather than compounding gains, and adopting more of them is not itself the win. Choosing and implementing the right one is ⁵. RICS says the same from the construction side: uptake has plateaued, and the difficulty in realising improvements is probably compounded by the fragmented nature of the sector and unstandardised data ⁴.

It will not tell you whether the sector's problem is yours. ONS analysis shows construction output per hour reaching an index of 114.6 by 2020 against 1970, while the market sector reached 250, with multi-factor productivity roughly a third below its 1970 level ¹. That is context, not a diagnosis of your business.

It will not survive shallow adoption. ONS puts the growth in self-reported AI use, among businesses with ten or more staff, at roughly 12% to 35% since late 2023, but extensive use sits at just 10% of adopters and the average firm's tool count has barely moved since the series started ². A licence is not a capability.

It will not replace a measured pilot. The audit says where to look. Proving it on real jobs, with a baseline first and a stop condition agreed, is covered in running a safe 30-day pilot and measuring automation in time rather than money.

What the prize actually is

Error, and error in construction is expensive.

The Get It Right Initiative's own figures put the cost of avoidable error at around 21% of project value, inside a wider range other studies put at 10% to 25%, worth £10bn to £25bn a year across the industry ⁸. That work is from 2015 and 2016 and rework-cost estimates vary widely in the literature, so hold the precision loosely.

Resist the comfortable reading of it. Look at what GIRI actually ranks as the causes: inadequate planning tops the list, with late design changes and poorly communicated or poorly coordinated design information close behind ⁸. Much of that is not design quality but design information failing to move cleanly, which is the same category of problem this audit maps, arriving from the other direction.

Run properly, the audit changes the question the business is asking. Not how do we get people to work harder when margins tighten, but what are we paying good people to do that a computer should already be doing, and what would we rather they were doing instead. That argument is set out in your best people should not be moving information around. If you would rather not run it alone, AI Metric does this with contractors and consultants as the first step of any engagement.