Most of what gets written about AI and buildings is either a sales pitch or a warning. Neither helps much when you are standing in a stairwell deciding whether to trust the thing in your pocket.
So here is the honest version. We build this for a living, and these are the limits we work around every day.
What it is genuinely good at
Modern vision models are strong at recognition and description. Point one at a wall and it will reliably tell you there is cracking, roughly what pattern, and where it sits relative to the opening. It is good at reading what is written down: signage, switchboard labels, tags, plant plates, the date on a compliance sticker. It is very good at not getting bored, which matters more than it sounds. On the four-hundredth door it is exactly as attentive as it was on the first.
It is also good at structure. Give it a walkthrough and it will group what it saw by room, keep the sequence straight, and produce something shaped like the document you were going to write anyway.
That is a real capability, and it is worth having.
Where it goes blind
Scale, without a reference. A model cannot tell a 0.4 mm crack from a 4 mm crack from a photo. It has no ruler. It will happily describe “a fine crack” because fine is what that pattern usually looks like in its training data, and it will be wrong exactly when it matters. If the width is going to drive a recommendation, put a gauge, a coin, or your tape in the frame, or say the measurement out loud.
Anything behind a surface. Moisture in a wall, a cavity, the state of a tie, corrosion under a coating. If your eye cannot see it, the lens cannot see it either. A meter reading is data the model does not have unless you give it to it.
Cause. This is the big one. A model can see a stain. It cannot reliably tell you whether that stain is an active leak, a historic leak that has dried, condensation, or a failed seal three metres uphill. Staining looks like staining. Working out which one it is involves knowing the building, checking the weather, and often coming back. That is your job and it is not close to being automated.
Compliance. It can read a sign. It cannot tell you whether the door assembly actually achieves its rating, whether the penetration was sealed to the tested system, or whether the certificate on file matches what is in front of you. Anything that ends in a judgement against a standard is a professional call.
What is not there. The absent handrail, the missing damper, the smoke seal that was never fitted. Models are much better at describing what is in frame than at noticing what should be in frame and is not. Absence is an expert observation.
The pattern underneath
Nearly every one of those failures is the same failure: the model is working from appearance alone, and the thing you need to know is not visible in the appearance.
Which gives you a usable rule. If the answer is in the picture, the machine can help. If the answer is in your head, in your instrument, or behind the plasterboard, it cannot. Recognition and description sit on one side of that line. Diagnosis, measurement and compliance sit on the other.
What to do about it
The practical move is not to distrust the tool. It is to feed it the things it cannot get on its own.
- Say the reading. If you took a moisture reading, say the number aloud while you are pointing at the spot. Now it is in the record.
- Put a reference in frame whenever a dimension matters.
- Say the cause, not just the symptom. “Staining to the ceiling, I think this is the shower above, not the roof.” That sentence is worth more than the photo.
- Narrate the absence. “There is no handrail to this flight.” Nothing else will catch that.
- Keep the judgement calls yours. Ratings, causes, recommendations, and anything that ends up over your signature.
None of that requires buying anything. It is a capture habit, and it makes the record better whether a model ever touches it or not.
Why we are comfortable saying this
We would rather you knew the limits than found them out in a report.
Sully treats the walkthrough and the narration as one thing precisely because the narration carries what the footage cannot: the reading, the cause, the absence, the call. And it drafts. It does not sign. Every line traces back to the frame and the words behind it, so the checking is fast and you stay the author of record.
The line between what the machine sees and what you know is not a temporary gap that a bigger model closes next year. Some of it is physics. Build your process around it.