AI for construction
Four days. That is the usual distance between the site knowing the drawing does not match the steel and anyone in the office being able to act on it, and almost every recoverable cost on a job lives inside it. AI for construction is illustrated with autonomous plant and drone surveys; the money is in closing those four days.
Four days is the number
Almost every avoidable cost on a job traces to information that existed on site and had not reached anyone who could act on it. Shorten that lag and most of the rest follows.
- Variations recorded late, or not at all. Agreed verbally, remembered differently, contested at valuation. It is the single largest recoverable sum on most jobs and it is lost to a lack of paperwork rather than a lack of entitlement.
- Estimating from a blank page. Every tender starts again, even though the last twelve jobs contain most of the answer. Takeoff and pricing is skilled work being spent on retyping.
- Site diaries and daywork sheets that never get typed. Written in a van, photographed, and filed in a phone. When the claim needs them, they are unusable.
- Nobody knows which job is going wrong until the month-end. The signals — labour hours drifting, materials ordered twice, a subcontractor missing days — are all visible earlier, in different places, to different people.
Closing the four days
- Capture on site that takes seconds: a photo, a few spoken words, a location — turned into a dated, structured record attached to the right job and the right drawing revision.
- Variation packs assembled from those records: what was instructed, by whom, when, with the evidence attached, ready for the quantity surveyor rather than reconstructed from memory.
- A first-draft takeoff and price built from your own completed jobs and rates, for an estimator to correct instead of starting from nothing.
- A weekly job-health note: hours against programme, orders against budget, and the two or three things that changed direction this week.
What it will not touch
- It does not price a tender. It drafts from your history; a person owns the number that goes out, because a wrong tender price is a year of pain.
- It does not approve a variation or instruct work. It records what was said and hands it to whoever has authority.
- It makes no judgement about safety, structure or compliance. Those are qualified decisions and they stay with qualified people.
- It does not replace the drawing register. It attaches to it, and if there is no register that is the first job.
Three reasons to wait
- Your rates and completed jobs are not in any usable form. Getting them there is the first project and it has value on its own.
- Site staff will not use a phone for anything but calls. This lives or dies on capture taking under a minute; if it will not be used, do not build it.
- One job at a time, run by the owner, who is on site daily. The gap this closes is not open yet.
The parts it would be assembled from
- Photo factory — assembly at volume, already running: one input in, a full set of finished documents out — one real batch was 32 items in a single run. A variation pack is that machinery pointed at site records.
- Systems watchdog — the weekly-note discipline in its blunt form: something whose only job is to notice when a thing stopped happening, and say so the same day rather than at month end.
What contractors ask before committing
What does AI for construction realistically do today?
It shortens the distance between site and office. Capture on site in seconds, variation packs assembled with evidence attached, estimating drafted from your own history, and a weekly note on which job changed direction. Autonomous plant is a different industry with a different budget.
Will it replace our estimator?
No. It removes the retyping from estimating, not the judgement. An estimator correcting a draft built from your last twelve jobs prices more tenders per week, which is the actual constraint in most contractors.
How does the variation piece pay for itself?
Count the variations lost in the last year because the paperwork was not there at valuation. Most contractors know that number roughly and it is uncomfortable. The build cost is usually a fraction of one job's recovery.
Our site teams will not use software.
Then it has to be a photo and a voice note, and nothing else. If capture takes more than a minute it will not happen, and a system nobody uses is worse than the paper it replaced. That constraint shapes the whole build.
Does it work with our project management platform?
It attaches to it rather than competing with it. The drawing register, the programme and the cost model stay where they are; what gets built is the capture and the assembly around them.
Bring one job's variation history
Including the ones you did not recover. In the free hour I will tell you which of them would have been defensible with a dated record and a photo, what capture on site would have to look like for your teams to actually use it, and what the first build would cost to run.
Book that hour →