AI Architecture Lab
← [ back to papers ]
2026-08-25 · NOTE

Where AI actually works in AEC

1. Abstract

I have spent a lot of time on one question: why does AI so often fall short in AEC? Not whether it works, but which parts work, which parts do not, and where the ceiling is genuinely low. This is where I have landed so far.

2. Motivation

The question underneath is whether an AI-native practice is even possible. AEC runs on drawings, design and 3D, close to the worst case for a language model. But if AI is going to do real problem solving here, I want to know which part it is actually doing.

3. Probabilistic and deterministic

Most of the AI we use day to day is probabilistic. You ask ChatGPT or Claude or Gemini and get a slightly different answer every time, because what it is doing is picking the next word.

Deterministic is when the answer is already fixed and the system goes and gets it. If the answer lives in a spreadsheet, it does not think the answer up again. It reads the cell.

the same question, asked three times
What is the minimum clear width for a barrier-free door?
Probabilistic
composes an answer
  • 01About 850 mm.
  • 02850 mm, though it depends on the occupancy.
  • 03Typically 850 to 860 mm on a barrier-free route.
picks the next word, so the answer moves
Deterministic
fetches an answer
  • 01850 mm · OBC 3.8.3.3
  • 02850 mm · OBC 3.8.3.3
  • 03850 mm · OBC 3.8.3.3
reads a value that exists, so the answer holds
An illustration, not a logged run. By the third answer the left side has widened into a range, and a range is not something you can check a door against.

Probabilistic looks like an AI render. Deterministic looks like using AI to build a tool, and the parts inside it.

4. What fits which

Probabilistic
variety is the point
  • 01Concept
  • 02Render
  • 03Communicating an idea fast
  • 04Drawing
Deterministic
the same thing comes round again
  • 01Repetition
  • 02Fixed workflows
  • 03Everyday work
Which side a task belongs on comes down to whether you want a different answer next time.

Probabilistic feels like freestyle to me, and the models are good enough now that it genuinely helps me get to an idea. Deterministic feels like assembly: structuring a workflow, building a Revit add-on, working out the logic. AI does the middle of that and I put the pieces together.

5. Doing the deterministic side well

Feed it what the office already has, then go through the workflow looking for the repetition and the fixed steps. Those are the parts you can assemble with AI, and that is where the time comes back.

the next bottleneckFirm datapast projects · standards · templatesdecisions · QA rules · internal workflowRepeated workfixed, everydayOne vertical toolclear input and outputVerified assetreusable, hard to copy
One repeated problem at a time. Each finished vertical becomes the asset the next one starts from.

Over enough of them the separate tools start connecting into something that only fits this office.

6. Does this actually work? My thoughts

I think it comes down to how well the firm feeds it. When the data going in is uneven, or the scope is drawn too wide, it tends to break somewhere in the middle, because a wide scope gives the errors room to compound. So I would start narrow.

Which makes it matter who feeds it and who judges the result. Practices are busy, and this is real work on top of the work, so unless somebody owns it, that is what decides whether any of it lands.

And if a practice does end up AI-native, I think it gets there in small steps taken deep, not wide ones taken early. Going all the way down on one thing is what shows you where the scope should open next. The width arrives as a result, not as the plan.