Does AI actually work in construction?

The anxiety is real, and it is being sold to

Start with what the industry is feeling, because it explains the volume of marketing aimed at it. Almost three in five construction professionals now say they are anxious about their organisation falling behind without digital adoption — up from just over a third in 2023. NBS

That is the emotion the entire category is being sold on, this page’s author included. A page that reported only that figure would be doing the same thing it is about to criticise, so here is the part that complicates it.

The anxiety shows up in individuals more than in organisations. In a separate survey of over 2,200 professionals, 41% reported feeling overwhelmed by the pace of technological change and 22% were concerned about AI’s impact on their own role. RICS

Feeling behind and being behind are different problems with different solutions, and only one of them is solved by buying software.

What actually works today

The honest list is unglamorous. What works in construction right now is narrow, well-defined tasks where the input is already digital and a human checks the output.

• Back-office automation. Invoice coding, document classification, filling out repetitive forms from information the system already holds. Dull, and the clearest return of anything here.
• Estimating and takeoff. Reading quantities off drawings, matching line items to a cost catalogue. Still checked by an estimator, and faster than measuring by hand.
• Document and RFI search. Asking a question across thousands of pages of specifications, submittals and correspondence, and getting the paragraph rather than the file name.
• Photo documentation. Tagging, sorting and searching site photography, so a question about a condition eighteen months ago is answerable.
• Computer-vision safety review. Flagging missing PPE or an unsafe condition in site imagery, as a second pass rather than a replacement for supervision.

What does not work today is autonomous project management. Nothing on the market runs a job, sequences a schedule against real site conditions, or makes a commercial decision you would stand behind without someone senior reviewing it first. Vendors rarely claim this outright, but a good deal of marketing is shaped to let a buyer assume it.

Why implementations fail

The data problem comes first

A model is only as good as what it can read. The average construction project runs on spreadsheets, email threads, WhatsApp messages, PDFs and paper, and the relationships between those are held in people’s heads. An assistant pointed at that has nothing to work with. It is not that the answers are wrong; there is nothing to answer from.

This is the failure that gets reported as "the AI did not work", and it is the one nobody can sell you a fix for, because the fix is months of unglamorous work putting the job into a system before any model is involved.

Then the barriers that are not cost

Cost is the objection people expect, and it is not the top one. In a survey of 235 US general and trade contractors by Dodge Construction Network with CMiC — a construction software vendor, which is worth knowing when reading their research — the two most cited concerns were about the technology itself. Dodge / CMiC

Reliability leads, and it is the concern this page’s next section exists to address: most of it is answerable by asking a vendor better questions before buying. Data security is second, and it is a question about who sees your project data and what happens to it — which we have written about separately in who owns your construction data. Dodge / CMiC

Cost divides sharply by size: 49% of smaller firms called the cost of investing in AI a greater issue, against 26% of large firms. The same research found 86% of large contractors believe AI will give them a competitive advantage, against 69% of smaller and mid-sized ones. The firms least able to absorb a failed implementation are also the least convinced it will pay. Dodge / CMiC

And the field tools nobody tests properly

A tool that works in the office and not in a basement with no signal is a tool the field will stop using in a fortnight. Offline behaviour, one-handed use, and whether it works in gloves are not details — they decide whether the data that everything else depends on actually gets entered.

Nobody has published what it returns

Here is the thing this page could not find, after going looking for it specifically.

That absence is a finding rather than a gap in the searching. The largest of the surveys here, at more than 2,200 respondents, captures sentiment and adoption intentions rather than documented outcomes — it measures what people expect and fear, not what they got. Figures claiming measurable impact circulate widely in vendor and agency content, and the ones this page chased could not be traced to any primary research, so they are not quoted here. RICS

That is the single most useful thing to know before buying. An industry three years into a wave of adoption, with strong opinions in both directions, has not produced a public number for what the technology pays back. Anyone quoting one to you should be asked where it comes from.

How to evaluate an AI claim

These are the questions to put to any vendor selling AI, including this one. Most of the reliability concern that leads the table above is answerable here, before money changes hands.

1. What data does it read, and where does that data have to live?

   What a good answer sounds like: A specific list of what it can see, and an honest statement that it cannot see what is not in the system. Any vendor implying their assistant will make sense of your email and WhatsApp is describing something that does not exist.

2. Is this a model, or a rules engine with a new label?

   What a good answer sounds like: A direct answer. Rules engines are genuinely useful and much more predictable — there is nothing wrong with one. There is something wrong with selling one as AI, and a vendor who will not say which it is has told you which it is.

3. What happens when it is wrong?

   What a good answer sounds like: A described failure mode: it flags uncertainty, it shows its source, a human approves before anything is committed. "It is very accurate" is not an answer to this question.

4. Can I see the source for an answer it gives?

   What a good answer sounds like: Every answer traceable to the record it came from, so a wrong answer is diagnosable rather than mysterious. Without this you cannot tell a good answer from a confident one.

5. Who sees the data, and does it train anything?

   What a good answer sounds like: A named list of subprocessors, and a direct yes or no on whether your data trains their models, with a stated position on who owns improvements derived from it. Vagueness here is itself the answer.

6. Does it work on a phone, in the field, with no signal?

   What a good answer sounds like: A demonstration on a phone rather than a laptop, and a straight answer about offline behaviour. The data everything else depends on is entered in the field or it is not entered.

7. What did it cost the last three customers like me, and what did they get?

   What a good answer sounds like: Specifics, or an admission that they do not have them. Given no credible industry figure for returns exists, a vendor claiming a precise one should be asked for the study.

Where WorkHoist stands

The part that should stop some readers buying

The honest sequence is that the assistant is worth something after the jobs, costs and correspondence are in a system, not before. That is months of ordinary work with no AI in it, and it is the part nobody sells because it cannot be demonstrated in a sales call. If a reader takes one thing from this page, it should be that the order matters more than the choice of vendor.

What is actually true

WorkHoist exposes the whole system over MCP, an open protocol, on every plan — read and write, from any AI client the customer chooses. That is a different shape from a chat box bolted onto a product: the data is reachable from outside the software rather than only through a feature the vendor built and controls.

The read and write distinction matters and gets blurred in marketing, including ours. The in-app assistant is read-only: it answers questions about your jobs and cannot change anything. The write access is over MCP, where an outside client can create and update records with a credential the customer grants and can revoke.

On the evaluation questions above, the answers for WorkHoist are: it reads what is in WorkHoist and nothing else; it is a model rather than a rules engine; the in-app assistant sends data to OpenAI, whose published API policy is that data sent to it is not used to train their models; and MCP connections are off by default, administrator-enabled, and send your data to whichever provider the customer picked, under that provider’s terms rather than ours.

What WorkHoist does not have is a published figure for what its assistant returns, for the same reason nobody else does. This page is not going to invent one.

Questions contractors ask

Does AI actually work in construction?

AI works in construction for narrow, well-defined tasks where the input is already digital and a person checks the output — back-office automation, estimating and takeoff, searching documents and RFIs, photo tagging, and computer-vision safety review. It does not run projects autonomously. It also does not work at all where project information lives in spreadsheets, email threads, WhatsApp messages and PDFs rather than in a system, because a model cannot read what it cannot reach.

What is AI used for in construction today?

The uses with a real track record today are unglamorous: coding invoices and classifying documents, reading quantities off drawings for estimating and takeoff, answering questions across large volumes of specifications and correspondence, tagging and searching site photography, and flagging safety issues in site imagery as a second pass. Autonomous scheduling and project management are marketed but not delivered — no product on the market makes commercial decisions without senior review.

Why do construction AI projects fail?

Most construction AI projects fail because of data rather than the model. Project information typically sits across spreadsheets, email, messaging apps, PDFs and paper, with the relationships between them held in people’s heads, so an assistant pointed at it has nothing to work from. Beyond data, the most-cited concerns among contractors are reliability and accuracy of output at 57% and data security and privacy at 54% — both ahead of cost — along with field tools that do not work on a phone without signal.

Is AI worth it for a small contracting business?

It depends on whether the business’s work is already in a system. For a small contractor running jobs on spreadsheets and messaging apps, the answer is usually not yet: the first investment that pays is getting jobs, costs and correspondence into one place, which involves no AI at all. Cost weighs heavier on smaller firms — 49% call AI investment cost a greater issue against 26% of large firms — and smaller firms are also less convinced of the advantage, at 69% against 86%.

Will AI replace estimators or project managers?

Not on current evidence. The tasks AI performs reliably in construction today are components of those jobs rather than the jobs themselves — measuring quantities, classifying documents, searching records — and each is checked by the person whose name is on the number. Four in five construction organisations have not yet moved past an AI pilot, and 57% do not trust the accuracy of AI output, which is not the profile of a technology about to take over commercial decisions.

How do I tell real AI from marketing?

Ask what data the tool reads and where that data has to live; ask whether it is a model or a rules engine with a new label; ask what happens when it is wrong, and whether it shows the source for each answer; ask who sees the data and whether it trains anything; and ask whether it works on a phone in the field with no signal. A vendor quoting a precise figure for return on AI investment should be asked for the study, because no credible public figure currently exists.