From Field Measurements to an Estimate a Person Can Review: Why Preconstruction Is Where AI Pays First

Measurements sit until someone draws them up, prices sit until someone looks them up, and the customer waits on both. The time goes to the queue in front of one busy person, not to the arithmetic.

The person who runs the company's field operations put it in one line: the waste is waiting, not estimating. The arithmetic of an estimate is quick. The measurements sit until someone draws them up, the prices sit until someone looks them up, and the customer waits on both.

It is also the measure the field lead chose for the system: waiting time, rather than a count of estimates produced.

Opportunity, qualification, an approved site visit, scan and design with the trades, an estimate a person reviews, the customer's acceptance, and an approved handoff to operations. Each stage has an owner, and the approver changes with the job.

The build covers everything from a new inquiry to an accepted job handed to operations. A person approves at the points where money or a promise is at stake, and who that approver is depends on the job.

Each room is scanned with a phone at the site visit. The system reads that scan into rooms, walls and openings, and the quantities are computed by plain code, not guessed by a language model. A model is good at reading a messy note; it has no business doing square footage.

That reading step is one of two that already run on real data: it has read a real kitchen job's report, 18 rooms of it.

The quote math is only as good as the prices under it. Today the written price list covers 2 of 11 trades, so the other nine stay with a person until they are priced.

The quote math passes its tests. The problem is underneath it: the written price list covers 2 of the company's 11 trades. An AI that fills the other nine with confident guesses would produce a clean-looking estimate that nobody should sign.

So until each trade is priced in writing, its lines stay with a person. That is slower on paper and much faster than rebuilding a customer's trust after a bad number.

Each step is finished only when it can reproduce a real job the company already sold and built. We picked six, from a kitchen and a bathroom to a small commercial fit-out.

Here is the uncomfortable part. As of this writing, the whole chain has not run end to end on a real job. Pieces pass their tests; the chain has not. I found that when I checked, and it changed how the work is defined.

Every stage is now finished only when it can replay a real job the company already sold and built. We chose six of different sizes, from a kitchen and a bathroom to a small commercial fit-out. If a stage cannot reproduce what actually happened on those, it is not done, however good its demo looks.

Industry surveys put about a third of construction work hours into looking for information, conflict and rework. Much of that starts upstream, in a scope or estimate that was rushed or incomplete.

The industry numbers point the same way. FMI's survey work found construction teams spend about 35% of their work hours on non-optimal activities: looking for project information, resolving conflict and rework. McKinsey's research puts construction labor productivity growth at about 1 percent a year over two decades, against 2.8 percent for the world economy.

A good share of that rework starts upstream, in a scope that missed a room or a price that was a guess. Fix the handoff from site visit to estimate and every later stage starts cleaner.

Pricing judgment, design choices and the customer conversation stay human, and every estimate is approved before it leaves the building. And if your company does not yet write its trade prices down anywhere, start there, with a spreadsheet, before any AI. The system can only be as good as the list it reads.