Since GPT-6 Astra launched, the feed has filled with AI building houses in Blender and walking through them in Unreal. Integrators and building owners are asking us the obvious question: can AI do this now? The fear of missing out is real, and the tool is genuinely impressive.
At QA Graphics, we’re heavily invested in 3D production: Autodesk’s AEC Collection including Revit, 3ds Max, Cinema 4D, Octane Render and a large GPU render farm. We’re always looking for ways to work faster. When a new AI tool comes out, we don’t just watch the demos. We test it against the same source files our customers send us: outdated drawings, partial plans and conflicting sheets. If a tool can make us more efficient, we want to be among the first to use it.
So when GPT-6 Astra launched, we put it to work on a real BAS floor plan.
The test
The customer asked us to take an updated architectural PDF, model it in 3D to match a style example, and add the ductwork. We gave Astra the same job and the same files:
- Match the visual style of a 3D example from our library.
- Start with the overall Level 1 floor plan of a school.
- Add in the 1B addition, including its ductwork.
- Add the ductwork from the customer’s existing BAS graphics.
It’s a fair test, and a typical one. The real job isn’t drawing one plan. It’s reconciling several sources that were never drawn to agree with each other.
The inputs alone show how much judgment the work takes. The overall plan was marked “for reference” at 1/32″ scale, a navigation sheet rather than something to measure from. Two clouded areas were new construction layered onto an existing building. Its colored callouts were architectural room types, not HVAC zones. And there were no detailed mechanical drawings at all. The only record of the ductwork was the customer’s existing BAS graphics: stylized images already running in their system, at a different scale, that someone had to interpret before they could trust them.
What it got right
At a glance, Astra’s output looks convincing. The overall layout is recognizable: the gym, the auditorium, the storage addition and the classroom wings all sit roughly where they belong. And it produced a finished-looking render fast.
That’s exactly what makes it dangerous.
Where it failed as a deliverable
Put Astra’s output next to our delivered plan and the problems an operator would hit are easy to see. Our CAD manager summed it up: it keeps filling in information we never supplied.
The ductwork was invented
Astra added ductwork we never asked for, stamping the same trunk-and-diffuser pattern into nearly every classroom, including areas none of its inputs covered. In the real building, every wing runs differently: trunks follow corridors, branch where the rooms need air and connect back to specific mechanical spaces.
The equipment was invented
In the real building, the units sit together in a handful of mechanical rooms. Astra scattered equipment across dozens of rooms. To an operator, that’s equipment that doesn’t exist, in places nobody would look for it.
The rooms didn’t match
Astra added rooms that aren’t in the building and drew others out of proportion, turning classroom wings into neat, symmetric grids that don’t match the drawings. An operator looking for a specific room number wouldn’t find it.
Large parts of the core were left as blank boxes.
The worst part is that Astra’s guesses look exactly as finished as the parts it was given. Where it had source material, like the gym, it came close. Where it didn’t, it made something up. There’s no visual difference between what Astra knew and what it guessed, so an operator has no way to tell which ductwork is real.
“You just needed a better prompt”
We tried that. We reran the test with more detailed instructions, and the second result was better. The gym and auditorium seating came through, and the style was closer to our example.
But our CAD manager’s review found the same core problem: it was still filling in information we never supplied. It added landscaping and sidewalks that only confuse the viewer, ductwork we didn’t ask for, and rooms that weren’t correct or proportional.
Better instructions help, but consider what a complete prompt would have to say: which rooms exist, which drawing to trust where they disagree, which areas are new construction, where every duct actually runs and where the equipment really sits.
Then there’s style. The customer wanted the new plan to match the existing look and feel of their system. Describing a visual style precisely enough for an AI to match it, down to line weights, colors, materials and how equipment is represented, is its own hard problem. A designer who knows the project sees it at a glance.
And close isn’t good enough. Astra’s style came near the example, but not consistently. BAS graphics get revised for years: a new VAV, a remodeled wing, a renamed room. Every update has to match the existing graphics pixel for pixel, or the system starts to look patched together and techs lose trust in what they’re seeing. A tool that lands slightly differently every time can’t maintain a graphics set.
By the time you’ve written all that down, you’ve done the hard part of the job yourself. The prompt just becomes a slower way to do it.
We found that out directly. We started correcting Astra’s output and stopped. Fixing invented ductwork, missing rooms and equipment that didn’t exist was more work than building it right in the first place.
What our team delivered
Our designers delivered the full job: two floors, with six area views on the first floor and three on the second.
Unlike Astra’s output, it’s correct, and getting it there took judgment no prompt supplied:
- Reading the existing graphics. The ductwork existed only in the customer’s stylized BAS graphics. The drafter had to work out what those images were actually representing before using any of it.
- Re-engineering without mechanical drawings. Using Revit, architectural modeling practice and years of BAS work, the drafter rebuilt a well-modeled duct layout from those graphics and the updated walls.
- Fitting the project’s new style. The plan had to match the existing style and look and feel the customer requested, while the customer was also updating the style of the overall project.
The result: ductwork that follows the building wing by wing, equipment in the mechanical rooms where a technician will find it, and rooms that match the plans. That’s what an operator needs at 2 a.m. when a space is too hot.
We asked an AI if it agreed
While writing this article in October 2026, we asked Claude, Anthropic’s AI assistant, whether it agreed with our findings. Its answer was candid:
“Models like Astra, and like me, are built to produce output that looks right. When the source material runs out, we don’t stop. We fill the gap with something plausible, and it comes out just as polished as the parts we actually knew.”
It also made the point we think will age best. Even a much better model still needs someone to decide which source to trust. AI working inside tools like Revit, with structured inputs, will likely improve faster than AI generating images from PDFs. But the bottleneck isn’t drawing. It’s knowing what’s true about the building.
Where this leaves AI
This isn’t a verdict that AI doesn’t work. It’s “not yet” for this kind of deliverable. We’ll keep testing every new release against real project files, and when the tools are ready, QA Graphics will be on the front line of putting them to work for our customers.
The value of a BAS graphic was never in generating the picture. It’s in knowing the picture is true: that the room, the duct and the unit on the screen are the ones in the building. That’s the part our team does, and it’s the part no demo shows.
