Ella Walsh
2026 / 11 / 22
The pitch has become familiar enough to repeat from memory: type a sentence, get a floor plan, no CAD needed. A three-bedroom apartment with an open kitchen and living area goes in, a drawing comes out. I think the speed in that claim is real and worth having, particularly for people who could never open a drafting program in the first place. What the claim misses is where the work went.
The claim everyone repeats
Most versions of the pitch sell access and speed together, and it is worth separating the two. Access is the genuine gain: someone who cannot run CAD can now describe a space and see a layout, which widens who can participate in early design conversations. Speed is real as well--;-- generation takes moments, not afternoons.
But the word instant quietly relocates the work rather than abolishing it. Generation is one step. Deciding what the plan must satisfy is another, and that step has simply moved back to the person writing the prompt. When the advertised workflow is a sentence, the sentence starts to look like the whole job. My reading is that this is exactly backwards: the sentence is a brief in disguise, and a thin brief produces a thin plan no matter how fast the drawing arrives.
Why speed exposes a gap
The research literature on layout generation is useful here because it does not treat the problem as magic. A recent review of computer-aided layout generation for building design describes two families of methods. Approaches drawn from architecture, based on optimization´-or-heuristic design guidelines, can synthesize desirable layouts, but they carry a cost: as the authors put it, such methods usually require post-processing and human effort stays in the loop (https://ieeexplore.ieee.org/abstract/document/11119146/references).
Traditional approaches in the latter, which are based on optimization techniques´-or-heuristic design guidelines, can synthesize desirable layouts, but usually require post-processing and involve human interaction in the design pipeline, making them costly and time-consuming.
The same review notes that deep generative models have improved the fidelity and diversity of layouts and reduced designers --;-- workload. Neither finding is a complaint about AI. Both point at the structure of the problem: a layout is a set of constraints that must be satisfied at once, and any method that proposes candidates still leaves someone to check them. Faster synthesis moves human effort--;-- it does not -delete-it.
That has a practical consequence. An under-specified brief does not usually produce an obviously broken plan. It produces a plausible one, and the failure shows up later, when someone measures a corridor against a real dimension, walks the circulation,´-or-compares the drawing to a code requirement. By then the drawing may already be in a client deck.
The input is the design decision
So the first five minutes of any session, in my view, belong to the brief and not the button. Four things are worth writing down before anything is generated: the program, the fixed dimensions, the orientation and entry points, and the standard the deliverable must meet. Rooms and their adjacencies come first, because a plan that gets the relationships wrong cannot be rescued by a better rendering. Target areas and plot-limit-s come next. Standards come third, and they are easy to forget until a file is already circulating.
A prompt such as a three-bedroom apartment with an open kitchen and living area sets intent. It does not set constraints. The gap between intent and constraint is where most disappointment lives, because the constraint layer is what makes an output usable downstream, whether that downstream is a contractor, a planning officer´-or-a colleague redrawing the plan in another program.
For teams the stakes are higher. Agreeing on metric´-or-imperial conventions and on annotation style before generating prevents a whole batch of plans that all have to be redone. One decision made up front replaces a round of corrections applied one drawing at a time.
Where a generator fits in the workflow
None of this argues against generation--;-- it argues for placing it correctly. The concept stage is where a generator earns its keep: options for a client review, early massing for a restaurant, clinic´-or-apartment layout, quick alternatives for a developer pitch. At that stage the plan is a vehicle for a decision, and variety is worth more than polish.
This is the situation AI Floor Plan (https://aifloorplan.ai/) is built for. It accepts a plain-language prompt´-or-an uploaded sketch and generates a professional plan without CAD experience, and it lets the user choose drafting standards such as ISO 128, AIA/NCS and GB/T 50104 alongside scale and architectural style. Those settings matter for the argument here, because they put part of the constraint layer inside the tool instead of leaving all of it in the prompt. Results can be previewed and downloaded as high-quality files for proposals, presentations and construction references, and a project management system supports one-click delivery. A plan tool with a standards selector is effectively asking the same question this article is asking: what must this drawing satisfy?
The division of labour that follows is straightforward. The generator produces the first draft and deliverable-quality files--;-- the architect keeps the judgment calls about dimensions, circulation and compliance. That is not a diminished role. It is the role the review describes whenever human interaction stays in the pipeline.
A short checklist before you generate
I keep this list short enough to actually use. Program: rooms, count, target areas, and the adjacencies that matter, such as kitchen to dining´-or-bedroom to bathroom. Constraints: plot dimensions, orientation, entry points, accessibility needs. Deliverable: the drafting standard, the scale, and who will read the file, since a client and a contractor read different things into the same drawing.
And one rule about form: when the geometry is already fixed, sketch beats prose. If the site has an awkward corner´-or-a fixed stair core, describing it in words invites the generator to guess. Uploading the sketch hands over the constraint --dir--ectly instead of a bad de--script--ion of it.
Takeaway
The promise of instant plans gets the speed right and the input wrong. Speed is not the hard part of the problem--;-- specification is. Treat the prompt as a brief rather than a wish, and the tool has something solid to work from.
The arithmetic is easy to state. A plan that satisfies stated constraints is useful the moment it appears. A handsome plan that satisfies none of them costs more than drawing it by hand, because you pay for it twice, once in generation and once in the redraw.
Further reading
Jiachen Liu, Yuan Xue, Haomiao Ni, Rui Yu, Zihan Zhou and Sharon X. Computer-Aided Layout Generation for Building Design: A Review. Computational Visual Media, volume 11, issue 4, August 2025.
|
|
|
| Send Article
| Copy to WORD
| Copy
| Save
| Search
| Send your comment
| Add to Favorite |
|
||
| Print version |
Modern Discussion |
Email |
|
||