Only about six percent of architects report using AI regularly in their practice, according to AIA data, a figure that sounds like a profession dragging its feet until you look at what the early adopters are doing. A handful of studios have moved past off-the-shelf image models and started training custom ones on their own built work. The output looks like their buildings. That shift changes the principal's job in ways the profession hasn't fully named.
Picture the first twenty massings for a new site landing before lunch, already rendered in the studio's own language. The principal has moved from drawing to choosing, and choosing well is a different skill than producing well.
Decide Whether a Custom Model Is Worth the Lift
A generic image model trained on the whole internet will give you something plausible on the first prompt. It will also give you something that could have come from any studio on earth. The reason to train on your own drawings, plans, sections, massings, and the back-catalog of built projects, is to get output that reads as yours rather than a stock set of forms with your name typed under them. A recent piece from e-architect on AI models trained on a firm's own drawings walks through what that pipeline looks like for a mid-size practice.
The trade-off is real. A custom model needs clean, labeled data, engineering help to fine-tune and host it, and a feedback loop that keeps it current as the studio's work evolves. For a practice doing two bespoke houses a year, that's overkill. For a studio pushing out competition entries, feasibility studies, and option sets every week, the math shifts fast.
Decide What the Model Is Allowed to Draw First
The temptation is to point the model at every stage of the project. Resist it. The strongest case for a studio-trained model sits at the front, where options are cheap and the question is still 'what could this site want to be?' Later phases, detailing, tectonic resolution, the stubborn work of making a building actually buildable, are where generative tools remain weakest and where the studio's craft still earns its keep.
So decide the boundary before anyone opens the laptop:
- Massing and parti studies. Twenty variants, generated against the brief and the site, in the studio's recognizable moves. This is where the model earns its keep.
- Material and envelope sketches. Useful for client conversations and mood-setting, less useful as the record drawing.
- Design development. Hand this back to the humans. The 3D resolution work is where current tools stumble hardest.
- Construction documentation. Nowhere near the model. The liability and precision requirements make this a bad bet, full stop.
Decide How to Beat the First-Option Bias
Here is the quiet danger of a model that produces twenty good-looking massings in an hour. The anchoring effect is the well-documented human habit of leaning too hard on the first piece of information in front of you. The first massing the model shows the principal becomes an anchor. The next nineteen get judged against it instead of against the brief.
There are concrete ways to blunt this. Review the brief and the site constraints on paper before opening the output. Have an associate shuffle the order of the variants so the principal doesn't see them in the sequence the model emitted.
Score against criteria written down in advance, program fit, daylight, structure, cost signal, rather than against gut reaction to image one. None of this is unusual; it's the discipline that keeps curation from collapsing into preference.
Decide How Many Options Actually Help
Twenty is a round number, not a useful one. The number of options a reviewer can usefully hold in comparison at once tends to be fairly small; past that, decision quality can suffer and the reviewer may start reducing on surface features rather than substance. If the model can produce two hundred massings as cheaply as twenty, the principal's job is to narrow before reviewing, not to admire the volume.
Build a filter stage into the pipeline. A junior architect or an associate culls the output against hard constraints, setbacks, program area, envelope, and brings forward a short list the principal can weigh against each other.
The model produces breadth. Someone else produces the shortlist. The principal chooses.
Decide What Authorship Means Now
There is a reasonable worry inside the profession that if the model drew the first massing, the architect didn't design the building. A more useful framing is that authorship was rarely only drafting. It was taste, judgment, and the hundred small corrections that bend a concept toward a workable building.
All of those still belong to the principal. The drawing hand is the part the model takes; the editorial hand is the part that matters more now, not less.
The practical implication for the studio is that the principal's calendar needs to shift. Less time producing options. More time writing the brief the model is trained against, defining the criteria it is scored against, and sitting with the shortlist long enough to pick the right one.
The job hasn't shrunk. It has sharpened.