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

How to Calculate the ROI of Construction AI

A finance director asks what the return is. The vendor's answer is a case study about a company in another country with a different cost base, and the consultant's answer is a percentage with no arithmetic behind it. Neither survives the first sensible question, which is usually some version of: saved compared to what, and would we have paid for that work anyway?

You calculate the return on construction AI from four numbers, and only one of them comes from the supplier. The hours a task takes now, measured rather than remembered. What an hour of the person doing it actually costs, built up from payroll rather than guessed. The full cost of the system including the time spent checking its output. And the level of adoption you honestly expect, which is never everybody from month one. This guide shows how to assemble those four into a figure that survives scrutiny, using UK sources a reader can check.

One thing this page will not do is tell you what AI saves. That number does not exist in any form worth quoting, and a guide that supplied one would be doing the thing it is warning you about.

Why pounds figures keep collapsing

There is a strong argument, made elsewhere on this site, that automation should be measured in time rather than money, because a money claim smuggles in a rate, an overhead position and a counterfactual, and a sceptic only has to reject one of the three.

That argument is right, and this guide does not contradict it. It answers the question that comes next. Your finance director will not approve a business case denominated in hours, because a budget is in pounds. So the sequence matters: measure in time, because time is observable, then convert to money in front of the reader with every assumption visible and adjustable. The conversion is the part that gets hidden, and hiding it is what makes the resulting figure worthless.

Everything below is therefore a method rather than a result. Where a number appears, it is an input chosen to demonstrate the arithmetic, and it is labelled.

Step one: the hours, observed rather than remembered

Take two ordinary weeks. Have the specific people who do the specific task note what it actually takes, and count the timestamps you already hold for anything with a start and an end, such as enquiry received against first reply, or completion date against invoice date, over the last 20 instances.

Two rules make the difference between a baseline and a fiction. Do it before anything is built, because afterwards the before is gone and a reconstruction is exactly as contestable as the vendor's case study. And measure the task rather than the person, and say so out loud, because the moment staff believe the stopwatch is aimed at them the notes turn into diplomacy.

What you want at the end is not "the team spends too long on this". It is: this task, done this many times a year, currently takes this many minutes each time.

Step two: what an hour actually costs

This is where most business cases quietly inflate, because people divide the salary by the hours and stop. The salary is not the cost.

Start from a defensible base. For the construction sector, the Office for National Statistics (ONS) Annual Survey of Hours and Earnings (ASHE) gives a median gross annual pay for full-time employees of £43,100 in April 2025, with median total paid hours of 40.0 per week ². By occupation, the same survey gives quantity surveyors £52,537, construction project managers £47,665 and construction and building trades supervisors £45,021 ³. Use your own payroll if you have it. These are for sense checking, not for substituting.

Then add what employing somebody costs on top, all of which is verifiable for the current tax year:

Two adjustments can move that materially and both are in the same source. The Employment Allowance for 2026/27 is £10,500, which for a smaller employer can reduce the National Insurance line to nil, and the Apprenticeship Levy adds 0.5 per cent on pay bills over £3 million ⁴. The CITB Levy is not charged at all where the total wage bill is under £150,000, and is halved between £150,000 and £499,999 ⁸. Check which of those apply before using the figure below.

On the assumptions above, and before any of those adjustments, that gives a fully loaded salary of £50,072, which is 1.16 times the base. Note what has just happened, because it is the most useful thing in this guide: the commonly used rule of thumb for fully loaded cost is 1.25 to 1.4 times salary, and the verifiable statutory components only get you to 1.16. The gap is holiday and sickness cover, employer's liability insurance, training, IT, plant, vehicles and workspace, and those come from your own accounts rather than from any published source. Anyone who quotes you 1.35 without being able to say what is inside it is guessing, and so is anyone who quotes 1.16 as though the statutory floor were the whole cost.

For an hourly figure, statutory annual leave is 5.6 weeks, or 28 days for a five day week ⁷. That leaves 46.4 weeks, and at 40 paid hours a week gives 1,856 hours, so £50,072 over 1,856 hours is a modelled £26.98 an hour. The unrounded figure of £26.978 is what the arithmetic below uses. State the assumption where the reader can see it: this combines an annual median with a weekly hours median from the same survey, it treats paid hours as though they were productive hours, which ONS does not, and it ignores sickness and non-chargeable time. Your real figure will be higher.

Two cautions on the ONS figures. They are provisional 2025 estimates ¹, and ASHE covers employees only: it does not cover the self-employed ². In an industry with a large Construction Industry Scheme (CIS) self-employed population, a model built purely on ASHE describes your employed staff and not everyone on your site.

Step three: the subtractions that decide it

Gross time value is the easy half. The net figure is where business cases are won and lost, and one line does most of the damage.

net annual benefit =
    annual time value
  + evidenced avoided costs
  + evidenced recovered value
  - software costs
  - implementation costs
  - support and maintenance costs
  - human review costs

Human review is the line that gets left out, and leaving it out is not a rounding error. A process that saves 20 minutes and needs five minutes of checking saves 15, not 20. An ROI case with no review line is describing an unreviewed system, which is a different and considerably worse proposition than the one being approved. It is also, in some settings, a professional obligation rather than a preference: the Royal Institution of Chartered Surveyors (RICS) professional standard on responsible use of artificial intelligence, in effect since 9 March 2026, requires its members and regulated firms to undertake randomised dip samples of outputs at regular intervals, and to document their decision about the reliability of an output in writing, where the AI system has a material impact on the delivery of a surveying service ¹⁰. Where that applies to your commercial team, the review time is not optional and belongs in the model.

Avoided costs need evidence rather than assertion. "We would otherwise have hired a document controller" counts only where that hire was actually planned and budgeted, and not where it is a hypothetical constructed to improve the number.

Recovered value is the most dangerous line here, and it needs stating carefully. Better records can improve the prospect of recovering an entitlement. They do not produce one. The outcome of any claim depends on the contract, the facts and the tribunal, and no software determines it, so recovered value must never be modelled as certain and is usually best left out of the case entirely.

Step four: run it at three levels of adoption

A model that assumes everyone uses the system perfectly from the first week is not credible, and a finance director will say so. Run three cases and lead with the worst one.

Lead with the pessimistic case, because if the project does not work there it does not work, and a director who is shown that first will trust the rest of the paper.

A worked example, illustrative

The following is an illustrative case, not a measured result and not a customer. The inputs are chosen to demonstrate the arithmetic and the outputs are modelled throughout.

A contractor automates the entry of delivery tickets into a materials register.

Gross annual time value is 4 minutes times 6,000 tickets, divided by 60, times £26.98, which is £10,791. Note the trap avoided there. The 6,000 tickets are the total across everyone who handles them, so there is no multiplication by the number of users. Multiplying by three people who share those same 6,000 tickets would count the same saving three times, and that is the most common error in an ROI case and the easiest for a sceptical finance director to spot.

At full adoption, the modelled net annual benefit is £10,791 less £2,400 software, less £600 support, less £1,079 of review time, giving £6,712, and a payback of 10.7 months on the £6,000 implementation cost.

Now run the three cases, with review time scaling with volume. Every figure below is modelled from the inputs above:

The pessimistic case loses money. That is the honest output of this illustration and it is more useful than a flattering one, because it tells you exactly what to do next: this project is marginal at 6,000 tickets a year, and it becomes sound either at higher volume, at a lower implementation cost, or not at all. A vendor who shows you only the 10.7 month figure has not lied about the arithmetic. They have simply shown you the case that assumes everything goes right.

What this model cannot tell you

Three limits, and they matter more than the formula.

It cannot tell you whether the time saved gets used for anything. An hour returned to a site manager who then goes home on time is a real improvement in a job that burns people out, and it is not a cash saving, and a model that treats it as one is overstating. Say which you are claiming.

It cannot establish that AI improves productivity in construction generally. The official measure of output per hour worked, in current prices for 2025, puts construction of buildings at £33.99 and specialised construction activities at £33.93 against a whole economy figure of £48.53 ⁹. It is tempting to read the gap as available headroom. That inference is not supported: output per hour is gross value added divided by hours from a household survey, it is not revenue per hour or a measure of how hard anyone works, and the high civil engineering figure of £63.89 mostly reflects capital intensity ⁹.

And it cannot substitute for your own accounts. Every figure above that comes from a published source is a national median or a statutory rate. The parts that decide your answer, your loading above the statutory floor, your volumes and your adoption, are yours.

Where to start

Pick one task, baseline it for a fortnight, and build the four numbers. If the pessimistic case works, the project is probably sound. If only the optimistic case works, you have a proposal rather than a business case, and the honest move is to say so in the paper rather than to discover it in month nine.

The companion question, which processes are worth baselining at all, is covered in the guide to processes ranked by evidence, and the evidence about where time actually goes is in where AI saves time in UK construction.

We run this model on every workflow automation engagement, and the principle that we never invent a figure is set out alongside the work itself in our case studies.