Maket has over a million users. Type a description of the house you want, select a lot shape, and the platform returns a floor plan in less time than it takes to sharpen a drafting pencil. Rooms appear where rooms should go, hallways connect them, and doors open in reasonable directions. At a glance the plan looks like something a designer produced, and at $20 a month it costs roughly what a designer charges for fifteen minutes of billable time.
It is also, in its current form, a sketch on a napkin that someone laminated and framed. Impressive framing does not change what is underneath. And the distance between a sketch and a permit-ready set of residential construction documents is not a gap that better algorithms will close next quarter. It is a chasm defined by building codes, liability law, engineering physics, and the accumulated professional knowledge that separates a picture of a house from a house that stands up, passes inspection, and doesn't burn down with people inside it.
What the Tools Actually Produce
I spent two weeks testing six AI floor plan generators against the same brief: a 2,200-square-foot single-family home, four bedrooms, two and a half baths, attached two-car garage, on a hypothetical 7,500-square-foot R-1 lot in a jurisdiction that enforces IRC 2021 without significant local amendments. Six platforms went through the test: Maket, ArkDesign.ai, Planner 5D, RoomSketcher, Homestyler, and ChatGPT 4o with a detailed prompt requesting dimensioned output.
Every tool produced a floor plan, and not one produced a buildable floor plan.
Maket came closest, with reasonable room proportions, a functional kitchen-to-dining relationship, and circulation paths that a family could actually navigate without turning sideways. But the master bedroom measured 118 square feet before the closet, the egress window in bedroom four was drawn at 30 inches wide with no sill height specified (IRC R310.2.1 requires 5.7 square feet of net clear opening with a maximum 44-inch sill height), and the staircase to the second floor violated the riser-tread relationship in R311.7 by roughly an inch and a half. ArkDesign positioned the water heater in an interior closet with no combustion air pathway noted. Homestyler generated a layout with the kitchen island blocking the required 36-inch clear passage to the rear exit. ChatGPT produced a text description and a rough diagram that placed three bedrooms on the second floor above a single-story garage, with no structural indication of how the floor system would span that distance.
These are not exotic failures but the ordinary, invisible code violations that a licensed architect catches during design development because an architect knows the IRC the way a surgeon knows anatomy: not as a reference manual to consult, but as an internalized spatial logic that shapes every decision before pencil touches paper.
The 7% Problem
I counted. The IRC 2021, Chapters 3 through 9, the sections that govern everything from minimum room dimensions to structural load paths to fire-resistant construction to emergency egress to mechanical ventilation, contains 42 distinct regulatory sections that a residential plan reviewer evaluates before stamping a permit application.
AI floor plan generators engage with three of them. R304 sets minimum room areas, and most generators hit the 70-square-foot bedroom minimum (though the master bedroom in my Maket test fell short). R305 requires 7-foot ceiling heights in habitable rooms, and every generator assumes standard 8-foot ceilings, which satisfies the requirement by default rather than by design. R303 mandates natural light and ventilation, and the better generators place windows in every habitable room, though none calculate the 8% glazing-to-floor-area ratio the code actually requires.
What about the other 39 sections? R302 governs fire-resistant construction: separation between dwelling units, garage-to-house fire barriers, flame spread ratings for interior finishes, and no generator addresses any of it. R310 specifies emergency escape and rescue openings: minimum clear dimensions, maximum sill heights, window well requirements for below-grade bedrooms. No generator checks compliance, and several produced layouts that clearly violate it. R311 covers means of egress in granular detail: stair geometry, landing dimensions, door widths, hallway widths, the specific relationship between riser height and tread depth that prevents people from falling down stairs. The generators produce stairs, but they do not produce code-compliant stairs. R313, the residential sprinkler requirement adopted by many jurisdictions, is invisible to every tool I tested.
William Cohen, AIA, summarized this on LinkedIn with unusual precision: "AI is the easiest 15% of the process. The other 85%, including zoning, structure, MEP coordination, and permits, is where projects succeed or fail." My testing suggests he is being generous. Fifteen percent would mean these tools handle the entirety of schematic design. They handle a fraction of it.
The Permit Deficiency Baseline
Before condemning AI floor plan generators for producing permit-deficient plans, it is worth acknowledging that human designers produce permit-deficient plans at an extraordinary rate. The City of Surrey, British Columbia, reported that 80% of residential permit applications contained significant zoning deficiencies, requiring an average of 1.6 resubmissions before approval. This is not a technology problem so much as a complexity problem. The IRC alone runs over 800 pages, and most jurisdictions layer local amendments, zoning overlays, HOA restrictions, and state energy codes on top of it.
AI tools that check compliance, rather than generate designs, show more promise on this specific front. CodeComply.AI runs submitted plans against IBC, IRC, ADA, FHA, and NFPA requirements and flags violations before a human reviewer sees the application. LA County launched an AI-powered pilot through Archistar for zoning review of fire-damaged R-1 homes in Altadena. Naples, Florida deployed CityView's AI plan review trained on Florida Building Code. These are review tools, not design tools, and the distinction matters: they read plans that humans or AI created and check them against code, the way a spell checker reads prose without knowing how to write a sentence.
Who These Tools Actually Serve
Homeowners exploring spatial possibilities before hiring an architect. That is the honest answer, and it is not a small market.
A client who walks into a first meeting with a floor plan generated by Maket, even a code-noncompliant one, communicates spatial preferences more clearly than a client who says "I want an open concept with lots of natural light." The AI plan becomes a conversation starter, a visual vocabulary that helps the architect understand what the client imagines when they close their eyes and picture their house. Maket's million-plus users are not, in the main, licensed professionals. They are people who want to see their ideas take shape before committing $40,000 to $75,000 in architectural fees.
For architects themselves, these tools offer rapid option generation during the earliest phase of schematic design, the phase where quantity of exploration matters more than precision of documentation. An Automation in Construction systematic review of 161 papers found that 68.94% of AI usage in architecture occurs during early design phases. That research confirms what the tools suggest: AI is a brainstorming partner, not a design professional, and the profession knows it.
A ServiceTitan 2026 industry report found that only 25% of residential contractors use AI meaningfully, and nearly half of all contractors surveyed expressed a lack of trust in AI-generated outputs. This is not Luddism; contractors bear direct liability for building what the documents specify. When the documents come from a tool that does not understand fire separation, structural loads, or egress requirements, the contractor's skepticism is not a market adoption problem. It is rational risk assessment.
If You Are Building a Home
Use these tools for what they are: sketchpads with autocomplete. Generate layouts, explore room relationships, see what 2,200 square feet feels like with four bedrooms versus three. Print the result and bring it to your architect as a starting point for conversation, not as a substitute for professional design services.
Do not bring an AI-generated floor plan to a contractor and ask for a bid. A contractor cannot price what a contractor cannot build from, and a floor plan without structural details, mechanical layouts, electrical plans, plumbing diagrams, and a code compliance review is a picture, not a plan. That difference will cost you months of delay at the permit counter and thousands of dollars in professional fees to produce the documentation the AI never attempted.
If your budget is tight, the $20-per-month AI tool plus a subsequent engagement with a licensed designer who can develop the concept into permit-ready documents is a legitimate strategy for reducing schematic design costs. But the emphasis belongs on "subsequent engagement." The AI finishes where the hard work begins.
Limitations of This Analysis
My IRC section count treats each numbered section (R301 through R339, plus Chapters 4 through 9 equivalents) as a single unit, which understates the actual regulatory surface area because sections like R311 (means of egress) contain dozens of subsections with independent requirements. The 7% figure is therefore a generous upper bound on AI coverage, not a precise measurement. Testing used free or standard-tier plans; enterprise tiers of some platforms may include features I did not evaluate. The 80% permit deficiency rate is from a single Canadian municipality and may not generalize to U.S. jurisdictions with different review processes, though anecdotal evidence from U.S. plan reviewers suggests comparable rates. I did not test ARCHITEChTURES or TestFit, which target multifamily rather than single-family residential, and whose code compliance capabilities may differ. A 2026 JAIST study found text-to-image architectural models achieved only 70.5% accuracy on basic vertical configuration tasks, suggesting the spatial reasoning limitations extend beyond floor plan generators to the broader category of AI-assisted architectural design.