A researcher at Ostim Technical University in Ankara asked three AI tools to draw floor plans for sustainable housing across five climate zones. She fed each tool detailed prompts specifying passive solar strategies, room adjacencies, and climate-responsive design principles. ChatGPT, Microsoft Copilot, and LookX produced 31 layouts between them. When Tuğce Çelik reconstructed those plans in AutoCAD and ran daylight simulations on equinox and solstice dates, only eight of the 31 were coherent enough to simulate at all. LookX was thrown out entirely for producing plans with what the study diplomatically called "insufficient architectural legibility." That is academic language for: the drawings were not buildings.
Of the eight survivors, not one consistently oriented living spaces toward the sun.
I keep returning to that number. Twenty-three floor plans discarded before anyone even asked whether the rooms would be comfortable to inhabit, because the rooms were not rooms yet, just rectangles arranged with the spatial logic of a parking diagram. And of the eight that survived, the fundamental question any first-year architecture student learns to ask on day one of studio, "Where does the light come from?", went unanswered.
What the Tools Actually Optimize
There is no shortage of AI floor plan generators in 2026. Maket.ai in Montreal has crossed one million registered users at $30 per month for its Pro tier. Finch3D in Malmö charges €49 per month and claims to reduce preliminary design time by 80 percent, while TestFit handles site-level optimization for multifamily developers and Snaptrude positions itself as a cloud-native BIM alternative. Each promises speed, and each delivers it: you specify room counts, target square footage, adjacency preferences, and the software generates multiple layout options in seconds.
Speed is genuine, and nobody disputes that. A 2026 Chaos and Architizer survey of nearly 800 architects found that 85 percent of AI users report efficiency gains, with 43 percent identifying concept and pre-design as the area of greatest impact. Finch3D's graph-based approach can iterate through hundreds of massing configurations faster than an architect can sharpen a pencil.
But look at what the tools actually measure: room count compliance, square footage targets, adjacency rules like bedroom near bathroom, code setback requirements, spatial efficiency ratios. These are the metrics the algorithms optimize against, and they are the easiest dimensions of architecture to quantify, which is precisely why they are the least important ones to a person who will live inside the result.
The Constraint Problem
Architects who have tested these tools keep hitting the same wall. You upload a drawing with a 10-foot ceiling and 24-inch-wide cabinets. Your 10-foot ceiling comes back as 12 feet and your 24-inch cabinets stretch to 36 inches, because the algorithm decided those proportions fit better. Snaptrude's CEO documented this pattern after talking to dozens of practitioners: AI tools treat locked dimensions as aesthetic suggestions rather than engineering constraints. A 10-foot ceiling is not just a number. It represents a commitment backed by code requirements, structural coordination, MEP clearances, and the budget your client already approved. Change it and you have invalidated six decisions downstream.
Only 27 percent of AEC firms currently use AI in their operations, according to a 2025 ASCE survey. Sixty-two percent of architecture professionals in a separate Chaos/Architizer study say these tools are not ready for production use. Not "waiting for the next version." Not ready. Architects are, as one practitioner put it, kissing a lot of frogs and watching them stay frogs.
The reason is not rendering speed or interface design. It is that an architect specifying a constraint has made a design decision, and the tool does not recognize it as one, whereas Excel never overrides a cell value because a formula respects its inputs and these AI tools see inputs as inspiration.
What a Floor Plan Actually Is
A floor plan is not a diagram of rooms but a prediction of daily life. The kitchen faces east because morning light makes breakfast feel different than fluorescent light does, and because a west-facing kitchen in Phoenix collects heat all afternoon that the HVAC system then fights until midnight. The hallway compresses before opening into the living room because spatial compression creates a sense of arrival that makes a 1,400-square-foot house feel generous. The master bedroom sits as far from the street as the lot allows, not because a setback calculator placed it there but because the sound of a garbage truck at 6 a.m. erases whatever the bedroom's square footage promised on paper.
Phil Bernstein and Vincent Guerrero of Yale's School of Architecture recently revisited Stanford Anderson's 1966 lecture on "problem worrying," the idea that architectural design is not a discrete algorithmic procedure but a process of allowing the solution to evolve as you explore the problem. Anderson made that argument at the dawn of architectural computation, and sixty years later it remains the sharpest critique of AI-generated floor plans anyone has offered: these tools solve, but architecture worries.
The Fee Question Nobody Asks
Custom home architectural design typically costs $15,000 to $75,000, representing 8 to 15 percent of construction cost. Schematic design, the phase where floor plans take shape, accounts for roughly 35 percent of the architect's total service fee according to AIA benchmarks. If AI tools compress schematic design time by 80 percent, as Finch3D claims, that implies a potential time savings of $4,200 to $21,000 on a typical engagement.
| Project Fee | Schematic Share (35%) | 80% AI Time Savings | Net Savings Potential |
|---|---|---|---|
| $15,000 | $5,250 | $4,200 | Up to 28% of total fee |
| $50,000 | $17,500 | $14,000 | Up to 28% of total fee |
| $75,000 | $26,250 | $21,000 | Up to 28% of total fee |
That table assumes the AI output is usable, which the Çelik study suggests it often is not, and that architects pass time savings to clients, which market dynamics do not require. More critically, it assumes the 35 percent of the fee allocated to schematic design was entirely spent on floor plan iteration, when in reality it includes site analysis, client interviews, code research, massing studies, and the kind of spatial intuition that comes from twenty years of walking through finished buildings and remembering what worked.
If you are hiring an architect for a custom home, the question is not whether they use AI tools. Sixty-four percent are already experimenting. The question is what role those tools play, because an architect who uses Finch3D to generate 200 massing options and then spends three weeks refining the best one around your family's actual morning routine is using the tool well, while an architect who hands you the AI's third-best output with adjusted furniture blocks is not designing your home but selecting it from a catalog.
Strongest Case For
This critique has a limit, and intellectual honesty requires stating it plainly: roughly 85 percent of new American homes are built without an architect involved, according to estimates derived from Census Bureau Survey of Construction data showing that most single-family starts use builder stock plans rather than custom design. Most production home buyers choose from a builder's portfolio of six to twelve stock plans, walk through a model home, and sign a contract. For those buyers, a $30-per-month AI tool that generates a competent layout, one that respects basic adjacencies and code requirements even if it ignores solar orientation, is a meaningful improvement over a stock plan designed in 2004 and unchanged since. Maket.ai's real competition is not a $50,000 custom design engagement but a builder's spec plan and a sales office.
That distinction matters enormously. An AI floor plan is a better starting point than no floor plan, and for a first-time buyer working with a production builder, the ability to say "I want the kitchen here and the office there" using a tool that visualizes the result in real time is genuine empowerment, even if the tool cannot tell them why one orientation will make the house comfortable and another will not.
What to Ask Your Architect
If you are building custom and paying for design, ask these questions before the schematic phase begins. They separate architects who use AI as a starting point from those who use it as a deliverable.
First: has the layout been tested for daylight on both solstice dates, not rendered with pretty light but actually simulated, because Velux Daylight Visualizer is free and if your architect has not run a simulation the orientation is a guess?
Second: which constraints in the plan are locked and which are flexible? If the answer is "all flexible," the architect is still sketching. If the answer is specific, with a ceiling height locked to structural coordination and window placement locked to the view you discussed, the design has progressed past what an AI can produce.
Third: walk the plan in your mind at 6 a.m. on a Tuesday and ask yourself where the coffee is and where the morning light falls, whether you can see the backyard from the kitchen and whether you can hear the garage door from the bedroom. No AI tool models these experiences because they are not geometric. They are biographical, the accumulated spatial preferences of a specific family, and translating those preferences into built form is the work that no algorithm has learned to do.
Limitations
The Çelik study used text-to-image diffusion models (ChatGPT, Copilot, LookX), not purpose-built architectural tools like Finch3D or Maket.ai, and those dedicated tools may perform better on dimensional accuracy and plan coherence. The 8-of-31 figure represents a specific methodology using climate-adaptive prompts, not a general benchmark for all AI floor plan generation. AIA fee benchmarks vary by region and firm size, and the 35 percent schematic allocation is an average across a range. The 80 percent time reduction claimed by Finch3D is a vendor figure without independent verification. The Chaos/Architizer 2026 survey reached 800 respondents, predominantly from smaller firms (60 percent under 20 people) in North America (50 percent), so adoption rates at large global practices may differ. No study has compared client satisfaction with AI-generated versus architect-designed floor plans over the life of occupancy.