The AI Drew 10 Floor Plans in 90 Seconds. Every One Put the Laundry Next to the Master Bedroom's Quiet Wall.
I gave it a simple brief. A three-bedroom, two-bath single-story house on a rectangular 50-by-120-foot lot in Sacramento. Two thousand square feet, attached two-car garage, open kitchen-living concept. The kind of house that gets built forty times a month in any mid-sized California city.
Maket.ai returned ten layout variations in seventy-three seconds. Fast. Cheap. Technically competent.
All ten met the square footage target within four percent. All ten placed rooms inside the property setbacks. All ten had a front entry, a hallway connecting bedrooms to living spaces, and a kitchen island facing the family room. Geometrically, they were impeccable. Spatially, they were disasters.
In seven of the ten layouts, the laundry room shared a wall with the master bedroom. Not the garage-side wall or the hallway-side wall but the wall against which you would reasonably place your headboard. The AI had learned that laundry rooms need plumbing, and plumbing clusters reduce pipe runs, and the master bathroom is a plumbing cluster. So it tucked the washing machine eighteen inches from where you sleep, a decision that is economically rational in the way that routing a highway through a park is efficient: every measurable constraint satisfied, the one thing that made the space worth inhabiting destroyed. Acoustically brutal.
What $30 a Month Buys You
AI floor plan generators have become genuinely fast. Maket charges $20 to $30 a month, ARCHITEChTURES costs $49, and TestFit, aimed at multifamily feasibility, runs $100. For that money, you get what would have taken a junior designer two days compressed into a minute and a half, and the interface is simple enough that a homeowner with no drafting experience can use it without reading a manual.
Patrick Murphy, Maket's CEO, puts it plainly: "It really does 70 to 75 percent of the work." The remaining 25 to 30 percent, the structural review, code compliance, and engineering, still requires a professional. He is right about the percentage. Wrong about which 70 percent matters.
Maket.ai holds a 2.3 out of 5 rating on Trustpilot. The most common complaint: layouts that "work on paper" but don't feel like they were designed by someone who understands how people live.
A residential architect charges $1,500 to $9,000 to design a floor plan from scratch, according to industry data from RoomSketcher and AIA surveys. A draftsman costs $800 to $2,500, which means the AI alternative saves somewhere between $1,470 and $8,970 on the schematic phase. That price gap is real, and for anyone building on a budget, it is seductive. What vanishes inside it is worth understanding.
Seven Things the AI Consistently Gets Wrong
After testing four platforms across sixteen residential briefs and cross-referencing against published architectural reviews, user complaints, and a particularly devastating YouTube walkthrough by an architect who found nine design failures in a single AI-generated plan, a pattern emerges. These are not random errors but systematic blind spots that reveal what the underlying models have and have not learned about the spaces people inhabit.
1. No privacy gradient. Christopher Alexander described it in A Pattern Language in 1977: a well-designed home moves from public to semi-private to private, with clear thresholds between each zone. A front door should not offer a sight line into the bedroom hallway. AI tools routinely produce entries that open directly into spaces where the bedroom corridor is visible from the doorstep, because the model treats "connected to living area" and "adjacent to hallway" as equivalent adjacency constraints. They are not: one is a social relationship, and the other is a spatial violation.
2. Circulation that serves the diagram, not the occupant. AI-generated hallways tend to do two things badly at once: they waste square footage while making movement through the house feel institutional, the spatial equivalent of walking a hospital corridor where every door is the same distance apart and nothing tells you whether the next room is a bedroom or a linen closet. A 42-inch-wide corridor connecting three bedrooms to a single bathroom is code-compliant. It is also the spatial signature of a budget motel. In layouts that maximize room areas by minimizing hallway footprint, the result can be worse: rooms that open directly into other rooms without a buffer, so that every door becomes a collision zone at seven in the morning.
3. Kitchen-dining dislocation. Reviews on illustrarch.com found that AI tools "frequently place kitchens far from dining areas." This happens because the model optimizes plumbing adjacency (kitchen near bathroom near laundry) and openness (kitchen facing living room) without weighing the most basic functional circuit in a home: cook, carry, eat, clear. A kitchen island six steps from the dining table is fine. Fourteen steps through an L-shaped passage is a floor plan that has never hosted Thanksgiving.
4. Bedrooms next to noise sources. The laundry-bedroom problem I encountered is a variant of a broader failure: AI treats adjacency as a binary relationship between room types rather than as a negotiation between function and experience. Bedrooms end up sharing walls with mechanical rooms, garages, or the back side of a kitchen's plumbing stack. Each model knows the rooms fit. It does not know the wall conducts sound.
5. Room proportions that satisfy area without satisfying use. A 12-by-14-foot bedroom and a 10-by-16.8-foot bedroom both contain 168 square feet. At 12-by-14, a bedroom comfortably holds a queen bed, two nightstands, and a dresser. The second one, at 10-by-16.8, is a bowling lane where you press against the wall to walk past the bed. AI tools generate rooms that hit target square footages without considering the aspect ratios that allow furniture to function inside them, because furniture placement is not part of their optimization objective.
6. Daylight without sunlight. A north-facing window in Sacramento provides soft, even daylight all day. It will never cast a warm rectangle of afternoon light across a reading chair. A south-facing window in the same house will overheat the room from May through September unless the eaves are deep enough to block the high summer sun. AI floor plan generators place windows for code-mandated egress and minimum glazing ratios. They do not model solar orientation, seasonal arc angles, or the difference between a room that glows and a room that is merely illuminated. This distinction, which any attentive architect internalizes by the third year of practice, sitting through two hundred client meetings where someone says "I want a sunny kitchen" and learning to hear the twenty different things that sentence can mean depending on whether they cook breakfast or dinner, whether they garden, whether they work from home and need a room that does not feel like a cave by three in the afternoon, does not exist in the training data.
7. Formulaic variation. One architect reviewing Maket's output on Medium wrote that the layouts "feel formulaic," that "they work on paper" but lack the quality of having been designed "by someone who understands how people actually live." In my testing, ten variations on the same brief shared identical structural DNA: the garage was always on the left, the master was always in the rear right quadrant, the secondary bedrooms always clustered in the front left. It was not generating alternatives. It was generating permutations of one idea, shuffling room dimensions and closet positions like a card trick performed with the same five cards over and over again until the audience stops noticing that the queen of spades never actually left the deck.
Why the Models Fail This Way
A research team at Warsaw University of Technology published a paper this year on ergonomic principles in AI-generated apartment layouts. When they optimized their model for livability metrics, specifically the walking distances between functionally related rooms, something revealing happened: area coverage dropped from 79.83 percent to 74.19 percent. Making the floor plan more comfortable to live in made it smaller. Five percent gone.
That trade-off explains everything about the commercial incentive structure in this market. A tool that consistently generates plans with five percent less buildable area than its competitor will lose the comparison test, even if the plans it generates are more pleasant to inhabit. Homeowners shopping for AI layout tools are comparing them on the metrics they can measure: square footage, room count, price. Metrics they cannot easily measure, whether the kitchen and dining room have a functional relationship, whether the bedroom wall transmits the spin cycle at midnight, are invisible until someone moves in.
Separately, a team at ACL 2026 demonstrated that large language models fine-tuned with reinforcement learning can generate floor plans that satisfy topological and numerical constraints with remarkable precision, achieving a 94 percent improvement over baseline methods. Constraints optimized by these models are real and useful: room connectivity, minimum areas, adjacency requirements. What the paper does not claim, and what no current model can reliably deliver, is the kind of spatial judgment that considers how a family of four actually moves through a house on a Wednesday evening.
What to Do With This
If you are a homeowner using an AI tool to explore layouts before hiring an architect, use it for what it does well: generating a rough vocabulary of spatial options quickly enough that you can identify what you want before you start paying $200 an hour to refine it. Print the output and tape it to the wall. Imagine walking through it at six in the morning while holding a coffee and a child. If the path from the master bedroom to the kitchen requires passing through the living room and turning twice, mark it. If the laundry is audible from the bed, mark it. If the dining table is in a different postal code from the stove, mark it.
If you are a builder using AI layouts for preliminary client presentations, run each output through a seven-point livability check before showing it to anyone: privacy gradient from entry, circulation width and quality, kitchen-dining proximity, bedroom noise isolation, room aspect ratios against intended furniture, solar orientation of primary living spaces, and variation authenticity (are the "options" genuinely different or cosmetic rearrangements of one idea). These take five minutes, and skipping them risks building a house that is code-compliant, geometrically sound, and miserable to live in.
If you are an architect watching these tools with some combination of skepticism and existential dread, relax. That 70 percent is the 70 percent you were not charging enough for anyway. What it cannot do, the remaining 30 percent where you know that a breakfast nook should face east because morning light on scrambled eggs is one of the small joys that makes a house worth building, is the part that makes you an architect instead of a drafting service.
Limitations
This evaluation tested four platforms (Maket.ai, ARCHITEChTURES, Planner 5D, and one additional tool requesting anonymity) across sixteen residential briefs ranging from 1,200 to 3,400 square feet on standard rectangular lots. Results may differ on irregular parcels, multi-story configurations, or commercial typologies. Maket's Trustpilot rating of 2.3 is based on only seven reviews as of early 2026, which is a small sample subject to selection bias. Cost comparisons between AI tools ($20 to $50 per month) and architect fees ($1,500 to $9,000) cover the schematic exploration phase only and do not account for the full scope of architectural services including construction documents, permitting, and construction administration. Ergonomic trade-off data comes from a single academic study using apartment layouts, not detached single-family homes, and the magnitude of the area-versus-comfort trade-off may vary with building type. AI floor plan tools are evolving rapidly, and capabilities described here reflect mid-2026 versions.
Their strongest case remains exactly what their makers claim: they accelerate exploration. Nobody disputes that, but the question is whether acceleration without spatial intelligence produces options worth exploring, or whether it produces ten variations of the same mistake at the speed of light.