- A text prompt became a Revit model. We asked the VitruAI Building Modeller to model the Eiffel Tower in an open Revit file, tier by tier. It built 89 elements across four tiers, from foundations at −2 m to the antenna tip at 330 m.
- A second prompt became a sheet set. We asked for drawings at 1:100 on sheets: a plan per tier and four facades. The agent produced 20 sheets, with tier plans A-101 to A-104 and facade sheets A-201 to A-216.
- It filled in what the prompt left open. We gave no sheet size, layout or title block. The agent set cut planes, cropped each view, split tall facades into parts, packed several parts onto one sheet and generated a title block from scratch.
- It checked its own work. It ran dry runs before building, read element counts and heights back from Revit after each tier, and listed every failed call with its fix.
- It was straight about its limits. It couldn’t open the reference link, so it said so and used published figures. And it told us the model is about LOD 300, not the LOD 400 we asked for, and what LOD 400 would take.
Modelling is rarely the slow part. Drawing it up is.
Levels. Views. Cut planes. Crops. Sheets. A title block. Then the shuffle to make a tall elevation fit a sheet it was never going to fit.
Sound familiar?
None of that is design. It’s setup. And across a practice it adds up: in many firms, 30–60% of billable hours go to work that designs nothing.
So we put the AI building modeller for Revit inside VitruAI through a test. We gave it a brief that is hard on purpose: the Eiffel Tower. A dense iron lattice, curved legs, four tiers and an antenna, with almost nothing orthogonal about it.
Then we asked it to draw the result.
Here’s what happened, start to finish.
What happened when we asked an AI building modeller for Revit to model the Eiffel Tower?
The short answer: two prompts produced a 3D Revit model and a 20-sheet drawing set at 1:100.
The first prompt went into the VitruAI chat, next to an open Revit file:
Can you generate a 3D model of the Eiffel Tower in the active document:
- LOD 400
- You will find more details about the facility here (Wikipedia link)
- Build elements upward tier by tier.
- Zoom to fit the active view each time a new tier is completed, to keep the tower in the middle of the screen.
That’s the whole brief. No family library, no dimensions table, no Dynamo graph.
The agent built the tower tier by tier and re-framed the view after each one, as asked. When it finished, it posted a summary of what it built, what it couldn’t do, and what to do next.
The second prompt was even shorter:
Can you create drawings with scale 1:100 and place them on sheets:
- Plans per tier.
- Four facades.
The result was a full sheet set inside the same Revit file.
Takeaway: You can brief this agent the way you’d brief a junior: plain language, a reference, and how you want it built.
How does the agent build a Revit model from a text prompt?
It works inside the open model through functions it writes for the task.
The VitruAI chat describes the agent this way: it “works in the open model through agentic functions: reads it, changes it, and writes the functions a task needs, kept in the project’s function library.”
In practice, the run went like this:
- It planned each tier before building it. Before Tier 2, for example, it set out the scope: “the pillars from 57.6 to 115.7 m, the mid-height tie girders, the second-floor deck, girders, railings and pavilions.”
- It tested before it built. Most writes ran first as a dry run that was then undone. Functions that had already worked without problems ran directly.
- It used Revit’s own tools where they existed. Zoom to fit ran through the existing zoom function, which frames the selected elements.
- It checked the result in Revit. After each tier, it read element counts and heights back from the model.
Here’s what that read-back showed:
| Tier | Height | What it contains | Elements |
|---|---|---|---|
| 1 | −2 to 57.6 m | 16 masonry piers, 4 curved pillars, 4 decorative arches, belt girders, a 70.6 m first-floor deck with a 36 m open centre, 4 glass pavilions, railings | 41 |
| 2 | 57.6 to 115.7 m | 4 pillars, tie girders at 86 m, belt girders, a 41.4 m second-floor deck, a restaurant and a kiosk pavilion, railings | 23 |
| 3 | 115.7 to 279.4 m | Pillars merging at 180 m into a single shaft, a central lift frame, the third-floor deck, a glazed gallery, an open upper deck, safety fences | 18 |
| 4 | 279.4 to 330 m | Lattice top section, lantern room, cupola, a beacon at 300 m, a 3-part antenna to 330 m | 7 |
Total: 89 elements, 3,397 solids, all Generic Models in the live Revit file.
Takeaway: The agent doesn’t ask you to trust it. It shows you what it built, in numbers it read back from Revit.
What did the agent do when it couldn’t follow the brief?
It told us, then worked around it.
Two things in the first prompt didn’t go to plan, and the agent flagged both in its summary.
The reference link. It couldn’t open the Wikipedia page, because it has no web access from Revit. So it said that plainly and used the tower’s widely published figures instead: a 125 m square base, floors at 57.6, 115.7 and 276.1 m, about 300 m to the top of the structure and 330 m with the antenna. Where no figure was available, such as member sizes or the exact curve of the legs, it chose values to fit the overall shape.
The level of detail. We asked for LOD 400. The agent’s summary opens with a warning: “This is not LOD 400. It’s about LOD 300.” It explained why. Real LOD 400 would mean modelling the tower’s individual iron parts, rivets, connection plates, stairs and lift machinery, and that needs the real structural drawings.
It also listed what you’d do next. Save the model, because it can’t save for you. Switch to a Realistic or Shaded view to see the materials. And to get closer to LOD 400, add lattice layers inside each pillar, stairs and lift tracks, which “needs a reference: drawings or the dimensions you want used.”
Takeaway: An agent that tells you where it fell short is one you can review. That’s the bar for letting it into a live project model.
What did the second prompt ask for, and what did the agent decide on its own?
The prompt asked for very little. The agent made the rest of the decisions a drafter would make.
Here’s how the brief compares with what came back:
| What the prompt said | What the agent did |
|---|---|
| Drawings at 1:100 | Every view set to 1:100 |
| Plans per tier | Created a level for each tier, then plans A-101 to A-104 (Tier 1 at +57.60 through the summit at +291.00) |
| Four facades | North, East, South and West elevations, each split into four parts: A-201 to A-216 |
| Place them on sheets | Created all 20 sheets and placed the viewports, starting with a dry run |
| Nothing about cut planes | Cut each plan 1.5 m above its deck (2 m above the summit level), so it cuts through the pavilions and gallery and shows the girders and pillars below |
| Nothing about crops | Cropped the plans to 760, 440, 210 and 140 mm square on paper, and cropped all 24 facade segments, for example 680×670 mm for the ground-to-first-floor halves |
| Nothing about a title block | Generated one from scratch: firm, project, drawing title, date, scale and sheet number |
Look at the left column again. Almost none of the right column was requested.
Takeaway: A short prompt works because the agent fills the gaps with standard drawing decisions, and you can see each one.
Why did the agent add levels nobody asked for?
Because a plan view can’t exist without a level.
Our engineer points this out in the walkthrough. He didn’t ask for levels per tier. The agent created them anyway, because Revit needs a level to host each floor plan.
It’s a small detail. It’s also the difference between a tool that follows instructions and one that understands the model it’s working in.
What happens when the Eiffel Tower doesn’t fit on a sheet?
The agent splits it, then packs the pieces.
At 1:100 the tower is far too tall for a single sheet. The agent didn’t change the scale to make it fit. It kept 1:100 and divided each facade into four parts, with titles such as “East Facade – Part 1 of 4 – Ground to +64.00, left half”.
Then it solved the packing problem.
On sheet A-208, the agent placed three facade parts side by side, instead of spreading them over three sheets. Our engineer picked this out as a good example of how the agent handled drawing packing on its own.
Takeaway: The agent kept the scale you asked for and changed the layout to suit it.
Want to see what this would look like on one of your own schemes? We’ll walk you through the Building Modeller on a real brief.
What went wrong in the drawing run?
Two calls failed. The agent fixed both and listed them under “Failed calls and what changed”:
- Building the title block failed because the template had elements Revit wouldn’t let it delete. The fix: delete only what Revit allows.
- Creating the sheets failed because sheet names can’t contain a colon. The fix: use “ – ” instead. Nothing was kept from either failed call.
One cosmetic issue made it into the final set: an extra rectangle inside the title block. Our engineer’s verdict: ask the agent to remove it with one follow-up prompt.
“I didn’t provide a lot of details about it. But anyway, it was solved pretty well.”
— VitruAI engineer, from the video walkthrough
Here’s why this matters. An agent that reports its failed calls gives you something to review. One that hides them gives you surprises later in the project.
Takeaway: Read the summary, fix the leftovers with a follow-up prompt, and move on.
How could you use an AI building modeller on real projects?
The Eiffel Tower is a stress test. Your projects are the point.
The Building Modeller turns a sketch or a described scheme into real Revit elements at the level of detail you choose. It’s one agent in the VitruAI agent layer, which also covers BIM QA/QC, MEP routing and cost plans from your own rates.
If you want to try it on a live scheme, here’s the approach that worked in this test:
- Write the prompt like a brief. Say what to model, at what level of detail, and how you want it built.
- Give it the reference as files or numbers. A sketch, drawings or the dimensions you want used. The agent works inside Revit, so a web link isn’t enough on its own.
- Ask for the drawings in a second prompt. Scale, which plans, which elevations. Let the agent handle crops, cut planes and layout.
- Read the summary. Check the read-back numbers, the failed calls, the fixes and the next steps it lists.
- Fix leftovers with a follow-up prompt. Small issues, like the extra rectangle here, take one more message.
Every conversation also shows which files the agent read and changed, so your team can review exactly what happened in the model.
FAQ
Can an AI agent really build a Revit model from a text prompt?
Yes. In this test, the VitruAI Building Modeller built the Eiffel Tower tier by tier from a written prompt: 89 elements from −2 m to 330 m. It works inside the open Revit file and creates real Revit elements.
Are these real Revit elements or an imported mesh?
Real Revit elements. In this run they were created as Generic Models in the live file, so you can select, schedule and edit them like anything else in the model.
Did the model reach LOD 400?
No, and the agent said so in its first line: “This is not LOD 400. It’s about LOD 300.” It explained that true LOD 400 needs the real structural drawings, and listed the next modelling steps.
What if the agent can’t access my reference?
It tells you. Here it couldn’t open the Wikipedia link from inside Revit, said so, and used published figures instead. For real projects, give it drawings, a sketch or the dimensions you want used.
Can it create views, sheets and title blocks as well as the model?
Yes. From one short prompt it created levels, plan and elevation views at 1:100, 20 sheets, and a title block generated from scratch.
What happens when a step fails?
The agent fixes the call and reports it. In the drawing run, two calls failed (title block template, colon in sheet names). Both were fixed and listed in the summary.
Can it work from a sketch instead of a written prompt?
Yes. The Building Modeller turns a sketch or a described scheme into real Revit elements at the level of detail you choose.
Does our model data have to leave our office?
No. VitruAI can run as a self-hosted Revit add-in, with on-premise deployment for firms that need their data to stay in-house.
The takeaway
We started with the hours that go into setting up drawings rather than designing them.
In this test, two prompts covered the model, the levels, the views, the crops, the sheets and the title block. The agent made the drafting decisions the brief left open, kept the scale we asked for, and told us plainly what it couldn’t do and what it fixed.
Now picture that on your next scheme instead of a Paris landmark.
What would you hand it first?
Run the Building Modeller on a brief from your own practice. Book a free 30-minute consultation and we’ll show you how it would work on one of your schemes.
Book a free consultation → Explore the agents
About the author
Ian Arden, Founder & AI Systems Architect, VitruAI. Ian founded VitruAI, the AI agent layer for the built world: specialist agents that model, check, coordinate and cost inside the live Revit model. Earlier, he built early AppAssure, later sold to Dell for around $130M. VitruAI is a spin-off of ADAIA, which has supported 270+ founders.