Before a single construction document gets drawn for a custom home, there is a phase that architects call schematic design and clients call agony. You sit across from someone billing $150 an hour and describe the house you think you want. You move the kitchen, widen the hallway, rearrange the bedrooms. You wonder whether the guest room should face the garden or the street, and whether that question matters enough to spend another hour on. It always does. Six weeks later you have a floor plan you can live with, a lighter bank account, and the uneasy suspicion that the optimal arrangement of your rooms was out there somewhere, unexplored, because exploring it would have cost another four thousand dollars.
For a $500,000 custom home with an architect billing 12 percent of construction cost for full services, total design fees run roughly $60,000. Schematic design consumes about 15 percent of that: $9,000, covering 40 to 80 hours of the most expensive per-insight labor in the entire project, spent almost entirely on spatial exploration. Moving rectangles. Testing adjacencies. Asking whether the living room wants morning light or evening light, and whether the answer changes if the lot faces south instead of west.
Three academic papers published in 2026 suggest that a machine can now do parts of this work in under a second.
What the Labs Built This Year
At the Mila research institute in Montréal, a team led by Lara and colleagues fine-tuned a large language model on real floor plans, then used reinforcement learning to teach it dimensional constraints. Tell it you want a three-bedroom house with a kitchen adjacent to the dining room, a master suite no smaller than 200 square feet, and a total footprint under 1,800 square feet, and it generates a layout that satisfies those constraints 94 percent more reliably than prior methods. That figure, published in work accepted to ACL 2026 (Findings), measures relative reduction in constraint violations against baseline generative approaches, not just raw accuracy. You type a sentence describing your house, and the model draws it.
Separately, a group at CVPR 2026 introduced HouseMind, the first multimodal architecture that unifies floor plan understanding, generation, and editing within a single framework. It uses a VQ-VAE to discretize room layouts into tokens, bridging the gap between symbolic reasoning about adjacency and the continuous geometry of walls, doors, and window placements. Hand it an existing plan and ask it to enlarge the kitchen. It redraws the layout while preserving the spatial relationships you did not mention because you assumed they were obvious. The model runs locally, without cloud GPU access, which means a mid-range laptop could host it once the weights are released.
And at the ANZAScA conference, Guo and colleagues demonstrated a conditional generative adversarial network that evaluates daylight distribution across a floor plan in 0.5 seconds, achieving a structural similarity index of 0.93 against physics-based Honeybee/Radiance simulation. That simulation typically costs $1,000 to $5,000 when an architect commissions it as an add-on service during schematic design. The cGAN replicates its output in the time it takes to blink.
Twenty Dollars a Month
Maket, the most visible commercial tool in this space, charges $20 per month for its Pro tier: 300 credits covering text-to-floor-plan generation, conversational editing, 3D visualization, and DXF/PDF export for handing to a licensed architect. Its CEO, Patrick Murphy, claims the tool handles "70 to 75 percent of the work" that would otherwise occur in early-stage design conversations. ARCHITEChTURES, a Barcelona-based competitor, charges $49 per month and adds parametric optimization for multifamily layouts. ARK, at $180 per month, layers in automated code compliance checks.
Run the arithmetic against that $9,000 exploration phase. A homeowner who spends six months in Maket before engaging an architect pays $120 total. Even if the tool only replaces half the exploration hours, the remaining $4,500 architect bill plus $120 in software costs undercuts the traditional $9,000 by 49 percent.
| Approach | Exploration Cost | Duration | Output Quality |
|---|---|---|---|
| Architect only (40-80 hrs at $150/hr) | $6,000-$12,000 | 4-8 weeks | Site-aware, code-informed, spatially nuanced |
| AI tool only (Maket Pro, 6 months) | $120 | Self-paced | Dimensionally accurate, site-unaware |
| AI exploration + architect refinement | $120 + $3,000-$4,500 | 6-10 weeks | Better-briefed client, fewer billable dead ends |
But that table hides something important. Cost and duration are measurable, but output quality is not, and the difference between "dimensionally accurate" and "site-aware, code-informed, spatially nuanced" is the difference between a floor plan that satisfies constraints and one that makes a home feel right when you walk through the front door at six in the evening with grocery bags in both hands and a child trailing behind you who needs to reach the bathroom without crossing the kitchen.
What the Machine Cannot See
Every model described above optimizes for stated constraints. Tell it 200 square feet for the master, it gives you 200 square feet. Tell it the kitchen should adjoin the dining room, it obliges. None of them asks why. None of them knows that your lot slopes seven degrees to the southeast, that the neighbor's second-story addition will block afternoon light from any room facing west within four years, that the oak tree you refused to cut has a root system that will crack any foundation poured within 15 feet of the trunk, or that your city's setback requirements on the north boundary effectively eliminate the bedroom wing you spent three hours designing in the AI tool last Tuesday.
An architect billing $150 an hour knows these things because they drove to your lot, walked the grade, checked the survey, and read the zoning code over coffee before your first meeting. That $9,000 exploration phase is not just the cost of moving rectangles. It is the cost of a trained spatial intelligence applying judgment that has no text-prompt equivalent: integrating site conditions, municipal code, structural constraints, solar orientation, drainage, neighbor context, and the thousand small decisions that make a house belong on its specific piece of earth rather than floating in the abstract dimensional space where AI floor plans live.
Maket's 70-to-75 percent figure comes from the CEO of the company selling the tool, and no independent study has verified it. HouseMind does not accept site data as input. Mila's model was trained on a dataset of existing floor plans divorced from the lots they were built on. The cGAN daylight tool evaluates illumination after a plan exists, meaning it can tell you a room is dark but cannot inform the generative process that should have oriented the room differently in the first place.
A Hybrid Worth Testing
None of this means the tools are useless. They are genuinely useful, and the research trajectory is steep. A homeowner who arrives at their first architect meeting with twelve AI-generated layout variations, annotated with preferences, having already discovered that a four-bedroom plan on a 1,600-square-foot footprint requires compromises they are not willing to make, will have a more productive first session than one who walks in cold. That productivity translates directly into fewer billable hours spent on spatial dead ends the architect would have explored anyway at $150 each.
If you are planning a custom home or major remodel in the next twelve months, try this: spend $20 on Maket Pro, generate 15 to 20 layout variations with your real constraints (room count, square footage, adjacencies), live with them for a month, annotate what you love and what feels wrong, then hand the entire package to your architect as a design brief. You are not replacing the architect. You are arriving with a vocabulary for what you want that would have cost you $3,000 to $5,000 to develop in billable hours. The architect still does the hard work of anchoring the plan to your site, your code, your soil, your light. But the conversation starts at mile three instead of mile zero.
What remains genuinely uncertain is whether that AI-assisted exploration produces better design outcomes or merely faster ones, whether the homeowner who explored 20 machine-generated variations develops real spatial intuition or false confidence in arrangements that looked good on screen but fail in three dimensions when the ceiling heights and sight lines are real. Nobody has studied this. No peer-reviewed paper compares design satisfaction, post-occupancy performance, or change-order rates between AI-briefed and traditionally briefed architectural engagements. Until that research exists, the 70-to-75 percent claim remains a vendor assertion, the $9,000 exploration phase remains the industry's prevailing method, and the honest answer is that both numbers are probably wrong in ways we have not yet measured.
Where This Leaves You
Spend the $120. Do the exploration yourself before you pay someone else to do it. Walk into your architect's office with opinions forged by interaction rather than imagination. But do not confuse the fluency of a machine that draws rooms in half a second with the judgment of a professional who has watched families live in the spaces they designed, who knows that the kitchen island your AI tool placed perfectly equidistant from the stove and the sink will become a barricade when two adults try to cook Thanksgiving dinner simultaneously, who understands that a hallway is not just a connection between rooms but a transition between the public life of a home and the private one, and who charges $150 an hour because that understanding took twenty years to build and cannot be downloaded.