AI in construction: what small firms get, and what they do not
Two sessions in, we stopped.
The job was taking measurements off a design drawing at a landscape construction firm. A PDF with a scale on it, and a team who can do it by hand fairly quickly. We fed it to the AI and it came back with questions. Then it came back with more.
By the time somebody had answered them all, doing the takeoff the old way would have been quicker. And the figures have to be exact, because a price goes out of the door on the back of them.
That is AI in construction as I have actually found it, and it is not the version you get if you go looking for it.
So the honest answer was to leave that job alone and do something else instead.
Who the guides are written for
I read the pages ranking above this one before writing it, which seemed only fair.
Oracle. Microsoft. Autodesk. Procore. ConstructConnect. Every one of them sells construction software, and every one of them is describing the software it sells.
Between them they cover drones surveying sites, cameras spotting defects, sensors predicting when plant will fail, and clash detection across a building model. All of it real. None of it aimed at a firm with twelve people, two vans and a bookkeeper who comes in on Thursdays.
The other thing those pages have in common is that not one of them cites a single number. A great deal about what AI could do. Nothing about how many construction firms are actually doing it.
What the official numbers say
The ONS asked. In its survey of more than 38,000 UK businesses in June, construction came out at 13% using at least one AI technology. Information and communication came out at 58%.
Construction is the laggard in the official data, by some distance.
You can read that two ways. Either the industry is behind and needs to catch up, which is the reading the software vendors would prefer. Or a lot of sensible firms have had a look, failed to see what it was for, and gone back to work.
I think it is mostly the second, and I am not sure that makes them wrong.
The job everybody wants automated first
It is nearly always the estimating. Takeoff, measurement, pricing up a drawing.
That is the right instinct. It is the most repetitive skilled work in the business, it happens under time pressure, and it is where a mistake costs the most.
It is also the job that beat us.
One distinction matters here, because it is the difference between what I tried and what the vendors sell.
Specialist takeoff software with measurement built into it is a different product from a general AI assistant reading a PDF. The specialist tools are built for drawings, and one of them advertises measuring up to five pages of plans in around thirty seconds.
What I can tell you about is the other thing. A general assistant, a real drawing, a real scale, and a real price at the end of it. It asked more than it answered.
Why it went wrong, and what the research says
There is decent evidence on this, and it is not especially flattering to anybody.
A Harvard Business School study of 758 consultants at BCG working with GPT-4 found they completed 12.2% more tasks and worked 25.1% faster. On a task chosen deliberately to sit outside what the tool could do, the same people scored 19 points worse than the group working without it.
That is the shape of the thing. Inside what it can handle it is genuinely quick. Outside, it is confidently wrong, and confidently wrong about a measurement that becomes a price is worse than slow.
The other study worth knowing is smaller and more uncomfortable. METR watched sixteen experienced open source developers work on large codebases in 2025. Using AI tools they were 19% slower. They believed they had been 20% faster.
Sixteen developers writing software is not a construction firm, and I would not stretch it further than it goes. The perception gap is the part to carry across.
The job that worked instead
We left the takeoff and did the cost catalogue.
Prices in the system were updated when somebody noticed one had moved and had half an hour spare. The supplier invoices were already landing in one place. Feeding a month of them in, checking them against the current catalogue, and flagging what had changed is a dull, bounded, checkable job.
It is not impressive. Nobody is demonstrating it at an exhibition. It also works, which the takeoff did not.
A consultant who tells you what not to automate is worth more than one who says yes to everything. That is an easy thing to write and a harder thing to do, because the takeoff was the job they actually wanted solved.
Pick the job that already has a right answer
The rule that came out of those two sessions is about checking, not about tools.
A good first job has three properties. It is repetitive enough to be worth doing. It is bounded enough that you could describe it in one sentence. And somebody in the office can tell in under a minute whether the output is right.
The cost catalogue passes all three. A price either matches the invoice or it does not.
Takeoff fails the third one badly. To know whether a measurement is right, you have to do the measurement, which is the work you were trying to avoid in the first place.
That is the test worth applying before you start anything. If checking the answer costs about as much as working it out yourself, you have not saved a thing.
What actually helps a firm of twelve
The wins in a small construction business are almost all admin, not building.
A tender pack of sixty pages, reduced to the deadlines, the mandatory requirements and anything that would rule you out. Then checked against the document, because it will occasionally miss one.
A set of scribbled site notes turned into something the office can read. A method statement drafted off last month's with the details changed. The chasing email to a subcontractor you have written forty times. A variation letter.
Then the paperwork nobody counts as work. Onboarding documents for a new subcontractor. The snagging list somebody photographed on a phone and now has to type up. An operation and maintenance manual assembled from twelve supplier PDFs. The insurance renewal letter that has to go to nine people.
None of it is the interesting part of the job. All of it happens in the evening, because the day is on site.
None of that needs specialist software either. It needs somebody in the office who knows how to ask for it properly.
What it should not be doing is measuring anything that turns into a price without a person checking it, signing off anything safety critical, or sending a quote on its own.
The software problem underneath
Some of what gets called an AI opportunity is not one.
One construction management platform cannot overwrite a cost catalogue when you import an updated version. To load a new one, you delete the existing lines first. One at a time.
In practice people export the lot, change it outside the system, delete everything, and import it fresh.
No amount of AI fixes that. The honest answer there is better software, or a workaround, and somebody should say so rather than sell you a model to paper over it.
What it costs
The general assistants run at around twenty pounds a person a month on the individual plans, with business versions costing more. Prices move, so check the vendor's own page rather than mine.
Specialist construction AI is a different market. It is sold per seat or per project, and the numbers are rarely published in a way you can compare.
If you are being shown one, ask to trial it on your own drawings rather than the demonstration set. A tool that performs beautifully on a clean sample and struggles with your architect's PDFs is a tool you will have stopped using by March.
What I would not buy yet
If nobody in the business has used a general assistant properly for a month, I would not buy a specialist construction AI product on top of it.
Not because the products are bad. Because you will not yet know what you want it to do, and you will be paying per seat to find out. The firms that get value from the specialist tools are the ones who already understand where the line sits between work a machine can take and work it cannot.
That order round is cheaper and it is also faster. A month of somebody using what you already pay for will tell you more about your own processes than a procurement exercise will.
What you could do on Monday
Take the last tender document you were sent. Not a live one. An old one, where you already know the answers, so you can mark the work.
Ask for the submission deadline, the insurance requirements, and anything that would disqualify you. Then check those three things against the document yourself.
If it gets them right, you have learned where the line is for ten minutes of effort. If it misses one, you have learned something more useful, and it cost you an evening rather than a bid.
Either way, note how long that job normally takes you. You will want that number later.
The honest limit
AI is not going to price your jobs. On the evidence, and on two sessions of trying it properly with people who know what they are doing, it is not reliably going to measure them either.
What it will take off you is the paperwork around the work, and usually from the person least able to spare the evening. That is a smaller claim than the one on the exhibition stand, and it is the one I can stand behind.
If your cost catalogue is out of date, that is where I would start.