How construction companies actually use AI

Every vendor in this industry now has an AI page. Most of it is not shipping. These are the questions a contractor actually asks, and the honest answers.

Who is using it, really? Contractors more than designers. The design side has had computation for years. The site side is where the manual work is, and manual work is what AI removes first. The buyer is usually a project controls or planning lead, not a modeller.

What works today, without qualification? Three things. Classifying elements that arrived badly typed, which is nearly every real model. Clustering clash results so somebody reviews forty groups instead of four thousand items. And searching a project's documents by meaning rather than filename.

What is close but not there? Progress measured from site imagery. The technology works in good conditions and struggles with occlusion, weather, and the gap between "looks built" and "is signed off". It is close enough to be worth designing for, and not close enough to be trusted unsupervised.

What is still a demo? Generating schedules or designs. Both look convincing in a controlled example and fall over on the constraints that actually govern a site: access, crane time, trade sequencing, and the fact that the programme is a negotiated document rather than an optimisation problem.

Why does any of it make a client's work faster? Because the bottleneck was never the drawing. It is data entry and reconciliation: somebody typing progress in, somebody else reconciling it with what the model says, and a meeting to agree which of the two is right. Every hour AI saves in this industry is an hour of somebody re-entering information that already existed somewhere.

Where should a company start? Where they already have ground truth. If you have cameras and a schedule, start with progress. If you have a document library, start with search. Starting with the most impressive feature rather than the best-instrumented one is how these projects stall.

What is the one design rule? It proposes, a person decides. Detection is never certain, and a system that silently marks work complete or blocks a task on its own will be switched off within a fortnight. Every AI feature in the product I design for is built as a suggestion with a human confirming, for exactly that reason.

This started as a section of my BIM viewer case study, where the same rule shapes every AI feature in the product.