Four days ago, Jensen Huang stood in front of six thousand people at NVIDIA's GTC conference in Taipei and watched an AI agent design a house. It took a site plan, concept sketches, and a mood board as inputs. Then it opened Rhino, modeled the terrain and setbacks, proposed building forms optimized for cost and comfort, generated interior layouts with walls and circulation paths taking shape in real time, caught its own mistake, fixed it, exported the model to Blender, and rendered photoreal images from multiple viewpoints under different lighting conditions. All of it ran locally on a single chip called RTX Spark, which NVIDIA will ship in laptops this fall for roughly $2,000.
Applause. Six thousand people clapping, and nobody in the room asked who would stamp the drawings.
What $20 a Month Buys You
The tools already exist for considerably less than an NVIDIA keynote demo. Maket, an AI floor plan generator launched in 2023, charges $20 a month. That buys 300 credits, with each floor plan consuming 20, which means roughly fifteen complete residential layouts per month for the price of a mediocre lunch in any city where people actually hire architects. Describe what you want in plain language: a three-bedroom ranch with an open kitchen flowing into the living area, the primary suite on the opposite end of the house from the kids' rooms, a mudroom off the garage. Seconds later, you have editable floor plans.
What Maket added in early 2026 matters more than the generation itself. They call it "agentic editing," a conversational interface where you tell the AI to make the kitchen bigger, add a powder room near the entry, flip the layout. It modifies the plan in real time while you watch, iterating on room sizes and adjacencies through conversation rather than through a design interface. Patrick Murphy, Maket's CEO, told his own blog the tool "really does 70 to 75 percent of the work," and that homeowners reach a point where they meet an architect and save a fair amount of time and money within the process.
Seventy-five percent.
That number deserves scrutiny, because the 75 percent Maket handles and the 25 percent it does not are separated by a chasm that has nothing to do with technology and everything to do with law, liability, and the physical reality of standing water on a badly graded lot.
What the Remaining 25 Percent Costs
In every U.S. jurisdiction, residential construction documents require a licensed professional's review and signature before a building department will issue a permit. Rules vary by state and locality. Some states exempt small residential projects from requiring an architect's stamp if a licensed contractor is pulling the permit. Others require a registered design professional's seal on any structure over a certain square footage or number of stories. But the principle is universal: someone qualified has to verify that the design meets local building codes, structural requirements, egress standards, and accessibility rules before the first shovel touches dirt.
That someone is expensive. And the gap between AI generation costs and professional review fees is where homeowners consistently underestimate their budgets. A full set of custom residential architectural services, encompassing schematic design through construction documents and permit coordination, runs between 5 and 15 percent of total construction cost. For a $500,000 custom home, the math lands at $25,000 to $75,000. Even a "stamp review," where an architect examines plans they did not draw and affixes their seal if the work meets code, costs $2,000 to $5,000 for a straightforward single-family residence, more if the plans need corrections.
AI tools like Maket are compressing the exploration phase, the iterative process of figuring out what you actually want your house to look like, from weeks into hours. That compression is genuinely valuable. But these tools are not compressing the professional review, the code compliance check, the structural coordination, or the moment when a licensed architect puts their name, their insurance policy, and their career on a set of drawings and says: this is safe to build.
Who Pays When the AI Gets It Wrong
On June 3, three days ago, a company called UpCodes launched an AI-native plan review tool that automatically checks project drawings against 11 million locally adopted building code sections across 6,000 jurisdictions. It runs discipline-specific analyses, covering architecture, structural, MEP, fire protection, life safety, accessibility, and energy compliance. It flags issues by severity, links each one to the relevant drawing page and governing code section, and lets teams track resolution.
UpCodes has 800,000 AEC professionals using its platform for code research already, which gives it a distribution advantage that most compliance startups would need years to replicate. Its CEO, Scott Reynolds, framed the launch as transforming compliance review "from a one-person bottleneck into a team-wide capability." That framing is revealing. Even UpCodes, which is building the AI that checks building codes, is positioning the tool as a first pass that augments professional review. Not a replacement. Not yet.
There is a reason for that caution. It lives in the thicket of contract law, professional liability, and clickwrap agreements that most homeowners never read until something goes wrong with their foundation. Bilzin Sumberg, a law firm specializing in construction, published a detailed analysis of AI liability in homebuilding that identifies the central problem with uncomfortable clarity. Most AI software ships with clickwrap agreements that cap the vendor's liability at the cost of the subscription, Twenty dollars. That is Maket's contractual exposure if their AI generates a floor plan with a code violation that survives into construction and causes a $200,000 structural defect. The architect who stamped the plans, assuming they reviewed them carefully, remains liable under their professional obligations and insurance policy.
No appellate court has ruled on a case involving AI-generated architectural plans that caused a construction defect, so the entire liability question remains hypothetical until someone's roof leaks and the ensuing lawsuit forces a judge to decide whether the architect or the algorithm bears responsibility for the flashing detail that failed. Pure theory. Untested by litigation. But the American Institute of Architects' position is clear, blunt, and on the record, their October 2025 guidance stating it plainly: "Architects remain accountable for all work products, decisions, and representations made to clients, agencies, and the public, including work produced with AI assistance."
Including.
California's Business and Professions Code Section 5536.25 adds a wrinkle. A licensed architect who stamps plans is not responsible for damage caused by subsequent unauthorized changes to those plans. But they are responsible for the design as submitted. Full stop. If an AI drew the plan and the architect's review missed a non-compliant egress path or a drainage problem that a more experienced eye would have caught, the liability rests with the stamp, not the algorithm.
What AI Design Actually Cannot Do
I have spent twenty years thinking about how buildings relate to the ground they sit on, to the light that enters them, to the bodies that move through them. Floor plan generation, the task these tools perform, is a spatial optimization problem: arrange rooms within a boundary, satisfy adjacency constraints, minimize circulation waste. An AI trained on thousands of residential floor plans can perform this optimization with extraordinary speed and reasonable competence.
What it cannot do is stand on a lot in the late afternoon and notice that the prevailing breeze comes from the southwest, that the neighbor's second-story addition will shadow your kitchen window by November, that the soil at the northwest corner holds water three days after rain because the grade drops two inches over fifteen feet in a direction the topographic survey did not capture at sufficient resolution. These are the observations that turn a functional floor plan into a house that works, that feels right when you walk through it seven years later and cannot articulate exactly why.
I am not making a sentimental argument but a professional one: what AI handles is the part of architectural design that can be reduced to rules and constraints. Rooms need minimum dimensions, hallways need minimum widths, and doors need to swing in specified directions. Solvable problems, all of them. But the remaining fraction, that 25 percent Murphy's CEO math leaves untouched, is judgment, experience, and the accumulated spatial intelligence of someone who has watched buildings fail and succeed across hundreds of projects. That judgment is exactly what the stamp represents. And it is exactly what an AI tool priced at $20 a month is not selling you.
The Strongest Case for AI Design
The counterargument deserves its full weight. Residential architecture has a serious accessibility problem that AI tools are beginning to address, and millions of Americans cannot afford $25,000 to $75,000 in design fees for a custom home; they should not have to choose between a tract house and no professional design input at all. AI floor plan tools democratize the exploration phase. A family building their first home can generate and iterate on layouts for a month at a cost of $20, arrive at a design conversation with an architect already knowing what they want, and potentially reduce the professional engagement from full-service to a targeted stamp review that costs a fraction of comprehensive architectural services.
That reduction is meaningful, because if AI exploration saves a homeowner even two rounds of schematic revision with an architect billing $150 an hour, the tool pays for itself in a single session. If it helps a family in a rural county where the nearest architect is ninety miles away articulate their spatial priorities before driving to that appointment, the tool has done genuine good.
Researchers at Frontiers of Architectural Research published a 2026 study showing that AI-generated building designs achieved 70.5 percent accuracy in vertical configuration tests, with architecture students rating output quality above 4 out of 5 on a Likert scale. A tool called Versur now lets architects upload building codes into a knowledge base, then have AI agents evaluate designs against those codes using retrieval-augmented generation, citing specific regulatory clauses. Compliance is becoming an evaluative layer inside the design workflow rather than a separate manual checklist performed after the fact. Progress, yes.
These are real capabilities. They will improve. But they do not answer the question: who signs?
What You Should Do With This Information
If you are planning to build a custom home and considering AI design tools, use them. Maket at $20 a month, RoomSketcher at $12 to $24 a month, and free tools like Sweet Home 3D are legitimate ways to explore layouts, test room sizes, and develop a clear sense of what you want before engaging a professional. They will save you time and probably money in the exploration phase, which is the phase where most custom home clients burn through hours of billable architectural time discovering what they actually want versus what they thought they wanted when they started the conversation.
Do not confuse the exploration phase with the construction document phase. Different work entirely. An AI floor plan generator produces a schematic layout, not a buildable set of plans. It does not include structural engineering, mechanical systems routing, electrical layouts, plumbing runs, or the dozens of details that a building department will require before issuing a permit. You will still need a licensed professional to review, modify, and stamp whatever the AI produced, so budget for that. For a stamp review of AI-generated plans on a straightforward single-family home, expect $2,000 to $5,000. For a home with any complexity at all, whether that means a hillside lot, an unusual structural system, or a local planning department with strong aesthetic opinions about what fits the neighborhood character, expect considerably more.
If you are an architect considering whether to stamp AI-generated plans brought to you by a client, understand that the AIA says your professional liability attaches to everything you sign, regardless of who or what drew the original lines. Your E&O insurance carrier may have opinions about this. Hear them before you accept the engagement.
And if you watched Jensen Huang's demo in Taipei and thought the hard part of building a house is now solved, consider this: the agent designed a beautiful building on stage. It did not pull a permit, negotiate with a planning department, coordinate with a structural engineer, or stand in front of a building inspector explaining why the header above the garage door meets the prescriptive span table for the specified lumber grade. Huang's demo ended where the expensive, slow, legally consequential work begins.
Limitations
This analysis draws on published pricing, vendor claims, and regulatory guidance, not on tracked outcomes from real residential projects that used AI-generated floor plans through permitting and construction. No comprehensive dataset exists for how many homes have been designed using AI tools and subsequently built. Maket's 70-to-75 percent claim is self-reported by its CEO and has not been independently validated. UpCodes' Plan Review launched three days ago and has no published track record. Existing professional responsibility doctrine and contract law analysis inform the liability framework discussed here; no appellate court has yet ruled on a case involving AI-generated architectural plans that caused a construction defect. Cost ranges for architectural services vary significantly by market, project complexity, and the architect's experience level.