Why Coherence Will Define AI in AEC

Coherence is the easiest thing in the world for a language model. It is the hardest thing in the world for construction. That single gap explains why so much AI arrives in the industry with confidence and leaves without impact.

Coherence, put simply, is whether something makes sense. A language model has one job at its core. Look at the words so far and predict the next one that fits. Coherence in that world is local and cheap. Does this token follow from the last one. Does the sentence hold together. Models are extraordinary at it because the entire test lives inside the text.

Construction offers no such luxury. Here coherence is physical, and it does not live in one place.

The clearest example right now is in the automated take-off category. Reading a drawing set and pulling quantities is one of the hottest applications of AI in construction, and it looks deceptively simple. It is not. A drawing set is not one document, it is dozens that have to agree. A member shown in plan has to reconcile with the same member in elevation, and in the section, and in the schedule. A call-out on one sheet points to a detail on another. General notes on the cover govern everything downstream. Read each sheet in isolation and the model returns information that looks confident and is quietly wrong. Getting a take-off right is not a reading problem. It is a coherence problem, spread across the whole drawing set.

And that is only coherence across documents. The harder version lives in the structure itself.

Take the problem we are solving at ConnectionAI, steel connections. A connection is a sum of parts. A beam, a shear tab, bolts, welds. AI can be trained to predict each part on its own, and the earliest version of the Steel Brain did exactly that. But parts that are each plausible can still add up to something that makes no sense in the real world. Crazy bolt patterns. A tab that does not fit. Before anything else, the output has to be a coherent connection.

Then it gets harder, because a connection never stands alone. It sits in a joint where forces have to balance and loads have to reach the ground. It sits among neighbours, other connections on the same beam that ought to resemble one another rather than solve the same condition five different ways. A connection can be perfect in isolation and wrong the moment the frame widens.

And the frame keeps widening. A project is thousands of these decisions that all have to agree with one another for anything to actually get fabricated and erected. The higher the climb, the less the answer lives in the connection and the more it lives in the context around it. Most AI never makes it past the first rung. It was trained to make things that fit locally, and construction punishes exactly that.

This is the part that took years to see clearly, and it is why the Steel Brain was built for structural steel from the ground up rather than borrowed from a chatbot. Getting a connection right is table stakes. Staying coherent as the frame scales up is the actual work.

But there is a further question the industry has not started asking.

Coherence does not stop at the project. Every business that touches a job carries its own. Its preferences, its standards, its accumulated way of doing things. Two fabricators handed identical drawings will build them differently, and both can be right. That is not noise. It is identity. An AI that cannot hold it will produce work that is technically valid and quietly foreign to the people who have to own it.

So the real question for any steel business is not whether AI can produce an answer. It is this. Do you understand your own coherence. And does your AI understand it too.

AI in AEC will be defined by coherence. The businesses that grasp their own first will be the ones that get the most out of everything that comes next.

‍