AI agents inside Revit?

There is a question I get more and more often since I started talking about what we built in Vault: can this be done in Revit? The honest answer is that I have not built it yet. And that is exactly why I want to write about it now, before I have something to show, because the difference between a vision and a product is something the AI world blurs far too easily.

I already walked this road once. Together with Avant Leap I built VaultPilot, an agent that operates Autodesk Vault from natural language. What we learned was not how to connect an LLM to an API, which is the easy part. It was how to keep a probabilistic system from writing the wrong things into production data: dry-run before every write, validation against the real Vault schema, read-back after writing to confirm that what the API says happened actually happened, hard limits for irreversible operations and a complete audit trail of every action.

The same anatomy, another product

When I study the Revit API I see a familiar anatomy: elements, parameters, transactions. A model with its own schema, operations that modify shared data and a high cost when something gets written wrong. It is exactly the kind of environment where an agent can read very well and write very badly.

That is why I believe the recipe travels almost whole. First, the agent proposes and the human authorizes: in BIM, where a model is touched by many hands and feeds sheets, schedules and coordination, an unsupervised automatic write is a risk nobody should accept. Second, validate against the real model: an LLM is perfectly capable of inventing a parameter that does not exist or a family that merely sounds plausible, and the only defense is checking every intention against what the model actually contains. Third, read back after writing: every write is a claim you have to prove, and the read-back is the proof. Fourth, hard limits: there are operations an agent should never be able to execute, by design, no matter how hard the user insists.

What changes in BIM

There are differences I take seriously. A Revit model is a shared working environment, with worksets and synchronizations, and that raises concurrency questions that a PDM with check-in and check-out keeps more contained. Revit's transactionality helps: operations are grouped and can be undone as a unit. But the blast radius of a change is different, because a badly written parameter can propagate to dozens of sheets without anyone noticing until printing day.

All of that is design work to be done, not a blocker. And it gets solved the same way it was solved in Vault: build the gate before you build the agent.

Vision first, real case later

What I will not do is sell this as if it already existed. I am starting to build in Revit and I decided to document the process in public: what works, what does not, and which lessons from the manufacturing world carry over to BIM. When there is a real case, I will tell it with data. In the meantime, this is what there is: a recipe proven in another product and the conviction that gates matter more than capabilities.

If you work in BIM and there is a Revit workflow you would hand to an agent tomorrow, I want to hear about it. The first real cases are always born from a conversation.