Estimating
There is a job in this industry that has not changed in thirty years. A drawing set, a highlighter, a tally counter, and a Sunday.
Somebody sits down with forty sheets and counts detectors. Or sockets. Or cameras, or outlets, or emergency lights. They mark each one as they go so they do not double-count, they keep a running tally per sheet, and at the end they add it up and hope.
It is one of the last genuinely manual tasks left in a trade that has otherwise moved on, and it is worth understanding why it has survived, before looking at what actually replaces it.
Because the stakes are high and the tolerance is zero.
A take-off drives your price. Under-count by four per cent on a job with three thousand devices and you have given away the margin before you started. Over-count and you lose the tender to somebody who counted properly. There is no comfortable direction to be wrong in.
So the instinct is to do it by hand, slowly, because at least then you know who to blame. Every estimator has been handed a count from somewhere else that turned out to be wrong, and after that happens once you do it yourself.
Fair enough. But the cost is not just the weekend. It is that quantification becomes a bottleneck. You cannot price as many jobs as you would like, because there are only so many weekends and only so many people who can be trusted to do it.
Three things, consistently.
Fatigue. Sheets one to five get counted carefully. Sheets thirty to forty get counted at eleven at night. Accuracy is not constant across a set and everybody knows it.
Revisions. The set gets reissued at REV C with changes on nine sheets. In theory you recount those nine. In practice, under deadline, somebody eyeballs the clouds and adjusts.
No audit trail. The output is a number in a cell. If somebody questions it three weeks later, there is no way to show how it was arrived at short of doing the whole thing again.
The idea is simple. Instead of counting instances of a symbol by eye, you show the software the symbol once, and it finds every occurrence of it across the whole set.
Worth being precise about how this works, because there is a lot of vague AI talk in construction software at the moment and most of it deserves scepticism.
It is pattern matching against the drawing. You select a region of a sheet containing the symbol you care about — a smoke detector, a double socket, a camera — and it scans the set for matches. It shows you what it found, highlighted, on the drawing. You look at the highlights, correct anything obviously wrong, and then commit the count.
It is not making a judgement about what the drawing means. It is doing the mechanical part of the work, which is the part that takes the weekend.
Any tool that gives you a number without showing its working is useless for pricing, because you cannot defend it.
The review step is not a nicety. It is the whole reason this is usable. Detections are shown on the drawing where they were found, so you can see at a glance whether it has picked up something it should not have, or missed a corner of a sheet with unusual line weights. You are still the one who decides the count is right. You are just not the one doing the counting.
In practice most of the review time goes on a handful of sheets where the drawing quality is poor, which are exactly the sheets a human would get wrong at eleven at night anyway.
The part people underestimate. When you train a symbol, it is saved to your organisation’s library.
Most contractors install a fairly narrow range of kit. The same detector head, the same range of sockets, the same three camera bodies, job after job. Once those symbols are in your library, the next drawing set that uses them needs no training at all — you point the tool at the set and go.
The first project pays for the setup. Every one after that is close to free.
A number in a spreadsheet is only half the value. The count should flow into the things that depend on it.
Feed it into a take-off list and it becomes a bill of quantities you can price or turn into a purchase order. Feed it into material pools and you have a stock figure that decrements as tasks claim kit, so mid-job you know what is left without a stocktake. Feed it into the task register and you have the work itself, ready to pin to the drawing.
That chain — count, order, install, claim, reconcile — is the same chain the manual weekend feeds into, just without the retyping at each step.
It will not read a specification and tell you what the client wants. It will not spot that the drawing is wrong. It will not price the job, and it will not tell you that the room schedule has changed since the drawings were issued.
It counts symbols, accurately, quickly, and shows you where it found them. That is a narrow claim, and it is deliberately narrow, because the honest version of what AI does well in this trade is narrow and mechanical rather than clever.
The best test is a drawing set you have already quantified by hand, where you know the right answer. Run it, compare, and see how close it lands and how long the review takes.
That tells you more than any demo, because you are checking it against your own number rather than ours.
More on symbol takeoff and the AI tools, and on where the counts end up in purchasing and materials.
Try it on a real job
Graftix puts every task on the actual drawing, then hangs the materials, certificates and sign-offs off that pin. Seven days, every feature, no card.
Read-only live environment · No sign-up required