What AI actually does with a construction drawing
There's a lot of noise about AI in construction at the moment and most of it is written by people who have never had to find a drainage detail with a gang on standby at 7am. I've spent the past two years working with these tools on real drawings from real jobs and built a drawing extraction engine which powers Resection's drawing analysis across multiple features. A word on what I mean by AI, because the term gets waved around loosely. Underneath nearly every product being sold to this industry right now sits the same technology, the large language models that power ChatGPT and Claude. So here's what's actually going on when that technology reads a drawing, in plain terms, including the parts the sales decks skip.
The first thing to understand is that the AI doesn't see a drawing the way you do. A PDF from a CAD system usually carries a hidden layer of text underneath the picture, every dimension, every label, every schedule entry, sitting there as machine-readable words and numbers. To the AI, a drawing isn't a picture at all. It's thousands of scraps of text scattered across a coordinate space and the whole game is pairing them up correctly: this label with that value, this level with that manhole, this note with the detail it belongs to. Get the pairing right and you can read an entire drawing set in minutes. Get it subtly wrong and you get something far more dangerous than an error, which is a wrong answer that looks exactly like a right one.
Building the engine meant testing that technology bare, on its own, before deciding what to build around it. Which means everything I'm about to describe is what you get if you point ChatGPT or similar at your own drawings today, because it's the same engine underneath, minus the machinery we ended up having to build. And what we found is that the hidden layer lies in ways that never show on paper. We found sheets that printed perfectly whose hidden text was stored in a scrambled order no human ever sees. Given one such schedule, the raw AI pinned only about half the values to the right rows, confidently. Our engine, which sorts and structures that layer before the AI gets involved, read all forty-odd correctly. On a live job that's the difference between a take-off you can trust and one somebody has to re-check line by line, which is to say no saving at all.
The quietest failure we found is the one that should worry anyone doing exactly that pointing. Values landing two or three rows adrift in a schedule, the level from one manhole reported against another. Nothing invented, everything plausible and it took an experienced engineer at the actual sheet to catch it. Build to that number and you find out in the ground, with an excavator stood waiting and an argument starting about who pays for the correction.
And when the raw AI fails, it rarely tells you. We watched identical requests on the same sheet return a full read one minute and an empty one the next, both presented as a job well done. On screen that empty answer reads as reassurance, nothing found here, when the truth is nothing was looked at. One pattern held across everything we checked though, and it's the finding that convinced me this technology can be made fit for construction at all: the failures were almost never fabrication. Nothing was made up. Things landed in the wrong slot. Misattribution is a problem you can measure and engineer down. A machine that invented plausible values would be unusable for this work full stop, and that machine is not what we found.
Which is why I've ended up confident about the radical difference between the raw AI and an engineered system, and it has nothing to do with how clever the AI is. The engine sorts the hidden layer before the AI is asked anything, does a large share of the reading deterministically so no AI is involved at all, forces every extracted fact to carry the exact spot on the exact sheet it came from and refuses out loud when it can't read something rather than papering over it. Every one of those disciplines exists because a specific failure demanded it. That's the honest shape of AI progress in this industry. Not a cleverer engine but a system that has been caught out enough times, by design, to have had the lying beaten out of it.
None of which makes the technology magic. The AI can't measure anything, so at present it won't spot on a 2D drawing that two bars clash or a duct doesn't fit through a beam, because reading words off a sheet and reasoning about objects in space are different problems and the second one is genuinely hard. A scanned drawing or a wet-signed sheet photographed on somebody's phone has no hidden text layer at all, which is rougher territory again. Anyone who tells you their AI handles all of that today is probably misleading you.
If you're being shown these tools at the moment, and most directors I know are, the test I'd apply is simple. Ask it something you already know the answer to, on your own documents. Then ask where it got the answer from. If it can always take you to the sheet and the spot, whether on a schedule or a dense GA, you're looking at something built the hard way. If it just sounds certain but occasionally can't answer the source question, you've met a very fast colleague who never checks anything and you already know how far to trust one of those.
The fear of being left behind by this stuff is real, I hear it from directors most weeks and it's mostly reasonable. But nobody in this industry needs to become an AI expert, any more than you needed to understand hydraulics to buy an excavator. What you need is the one habit above: never accept an answer about your own job without asking where it came from. You've spent a career pricing the difference between a subbie who shows you his build-up and one who gives you a number with a straight face. Judging these tools is the same judgement, applied to a new salesman. The technology is new but the way you should buy it isn't.
Details of examples changed where needed; the patterns are real.