Article
An LLM reviewing your building file? Yes — but not like that
The most seductive idea I saw this year in AEC research: hand a building code and a BIM model to a language model and have it answer complies / does not comply. The temptation is obvious — code review is slow, manual and expensive. My take: it will happen, but the one who signs will remain human for a long time, and that is how it should be.
The trust spectrum
Not all code-related tasks are equal. Here is where, today, I would trust an LLM verdict:
What convinces me — and what doesn’t
Convinces me: the LLM as a tireless junior reviewer
A good workflow doesn’t ask the model “does it comply?”. It asks: find the applicable clauses, extract the parameters from the BIM model, show the comparison and cite your sources. Every step is verifiable. The engineer reviews reasoning, not pages. That alone turns weeks into days — and leaves an audit trail, which is what code review is supposed to be.
Doesn’t convince me: the verdict as a black box
A building code is not just text: it is interpretation, local context and legal liability. A wrong “does not comply” costs a redesign; a wrong “complies” can cost lives. As long as a model can hallucinate a clause with full confidence, its word cannot be the last one. The right question is not “how often is it right?” — it is “who answers when it fails?”.
My bet
In five years, automated code pre-review will be as normal as spellcheck: nobody will brag about using it, but nobody will want to work without it. The final signature will still belong to an engineer — with fewer hours lost hunting clauses and more hours thinking about what the code doesn’t cover. The right tool doesn’t replace judgment: it clears the desk so judgment can work.