Can ChatGPT do a takeoff? Not one you can price

So, can ChatGPT actually do a takeoff?

You've dropped the Level 2 lighting plan into ChatGPT, asked how many downlights are on it, and got a number back before your coffee's gone cold. Tempting, right?

Short answer: not for a takeoff you can price. And look, ChatGPT isn't useless on drawings. It'll read a sheet, find the finishes legend and tell you what a code means. That's real, useful work.

What it can't give you is a quantity with its evidence attached. Think about what a number you actually carry needs. A sheet, a revision, a scale, and a shape someone else can look at. A chat reply carries none of them. Just a number.

And you're the one who answers for every quantity in a tender. Subbies push back on the ones that look wrong (which, fair enough, is their job). If you can't show the working for a number, it isn't safe to carry.

What do OpenAI and Anthropic say about it?

Both companies publish the limits of their models' vision, right there in their own docs. OpenAI's vision guide says: "The model may give approximate counts for objects in images." Then it adds: "The model struggles with tasks requiring precise spatial localization."

Anthropic's vision documentation says much the same about Claude, that it "can give approximate counts of objects in an image but might not always be precisely accurate, especially with large numbers of small objects."

Now picture a reflected ceiling plan. That's a large number of small objects, pretty much by definition. So's a hydraulic layout. So's the line marking in a carpark. Both companies are basically telling you the same thing: a model's reading of a picture is an estimate, not a measurement.

Here's where a typed count bites you

Let's run it. The figures are illustrative, and money's ex GST. Say you give a chat the Level 2 lighting plan and ask how many DL1 downlights it shows.

1. The reply says "approximately 140 DL1 downlights".
2. The plan shows 152.
3. At A$185 each, supplied and installed, the 12 missing fittings are 12 × A$185 = A$2,220.
4. The electrician's quote allows for 152. Your figure says 140, and nothing shows you which one's right.

Honestly, the A$2,220 is the smaller problem. The bigger one is finding the 12. How would you even do that? You'd have to count all 152 symbols yourself, and by then, well, you've done the takeoff anyway.

And here's the annoying part. A right answer wouldn't leave you any better off. Say the reply had come back with 152. You still couldn't trust it without doing that same count.

Same models, so why's yours any different?

Fair question. The Estimator+ takeoff uses models from the same two companies we just quoted. So why should its quantities get more trust than a chat reply's?

It comes down to where the number comes from. In a chat, the model looks at a picture and types a total. Counting, scaling, arithmetic, the lot, all in one step that nobody can see.

In Estimator+, the quantity is measured on the sheet itself. Every dot, wall run and outline is drawn on the drawing, at its printed scale and revision, and the number comes from those shapes.

So a count is the number of dots. A length comes from the line and the scale. There's no total for anyone to make up, because every figure has a shape behind it.

Do the models' limits just vanish? No, and we'd be kidding you if we said they did. They become visible instead. A dot on the wrong symbol, or a wall line sitting off its wall, is right there on the drawing where you can see it and fix it. And a symbol with no dot stands out next to all the ones that have one.

Back on that lighting plan. The AI's line says 152, and there are 152 dots on the sheet, each one on a symbol. Then the electrician's quote comes in allowing for 146. You don't recount the floor. You send the electrician the sheet with the dots on it. Much shorter conversation, that one.

Want the rest? AI takeoff software covers what else the Estimator+ takeoff does, from the schedules to the checking.

So what's ChatGPT actually good for in estimating?

Plenty, honestly. ChatGPT reads the text on a sheet well: notes, legends, schedules, title blocks. Ask it what a code means or what a detail's showing and you'll get a useful answer.

Claude, ChatGPT or Codex can also work inside the tender. They'll answer questions from its lines, quotes and drawings, file a subcontractor's quote from a PDF, and fill empty rates when you ask, each with its source.

Got a specification? Attach it, and the AI keeps what affects the price on this job, filed against the trades.

But ask a plain chat for the takeoff itself and you're back to a typed total. AI takeoff software covers what the takeoff does and what it leaves to you.

So here's what we'd suggest. Grab a lighting plan you've already counted, bring it to a demo, and check the AI's dots against your own count. You already know the answer, so it's a fair test.