Your City's AI Approved the Plans. The Foundation Cracked. Who Pays?

Honolulu's Department of Planning and Permitting used to spend 90 minutes reviewing a residential plan submission on a good day, with an experienced reviewer working a project that didn't have weird setback issues or nonconforming lot dimensions. Now it takes 20 minutes, because the city deployed Clariti's AI plan review tool and saw a 70% reduction in residential review times and 64% overall, a transformation so compelling that Denver signed on in May 2026 and communities across 26 states now use the platform. At the University of Florida, a patent-pending system called AutoReview.AI ingests all 800-plus pages of the Florida Building Code and checks compliance in 30 minutes flat, replacing reviews that used to consume weeks of staff time. Bakersfield processed 500 permits through Symbium's instant system in six months, and after the LA fires Governor Newsom deployed Archistar's AI e-check software across LA City and County for free.

Everyone wants to tell you about the speed, but speed is not the interesting story here.

What matters is what happens when the AI approves a plan that shouldn't have been approved, the contractor builds faithfully to that plan, and 14 months later the homeowner notices the foundation cracking because the footing depth violated Section R403.1.4 of the IRC and nobody, human or algorithmic, caught it. The question that no statute, no court ruling, and no insurance policy has answered is brutally simple: who pays for that foundation?

How Liability Worked Before the Algorithm Arrived

Before AI entered the permitting workflow, the liability chain for a missed code violation was messy but mapped, with every participant carrying insurance products tailored to their specific role in the chain. Your architect carried errors-and-omissions insurance covering design mistakes, and if the E&O carrier balked, the claim at least had a well-litigated framework of precedent to travel through. The municipality employed a plan reviewer who was a civil servant protected by governmental immunity for discretionary functions, which in practice meant you generally couldn't sue the city because a human reviewer exercised professional judgment and courts treated that judgment as a discretionary act shielded from liability. Your general contractor's commercial general liability policy covered workmanship defects that emerged during or after construction. If a foundation cracked, attorneys could work the chain by asking whether it was a design error falling under the architect's E&O, a construction defect covered by the contractor's CGL, or some combination that split exposure across multiple policies. Nobody loved this arrangement, but everyone understood it, and the insurance products existed to cover every link because actuaries had decades of loss data to price against.

Now replace the human plan reviewer with an algorithm, and watch every assumption in that chain collapse simultaneously.

Four Parties, Zero Accountability

When an AI reviews your plans and misses a code violation, four parties will argue with escalating creativity that the failure belongs to someone else.

AI vendors wrap their products in end-user license agreements containing standard SaaS terms that disclaim liability for errors in output, cap damages at the subscription fee paid, and state explicitly that the tool is "advisory only." Clariti, AutoReview.AI, Symbium, CivCheck, CodeComply.Ai: the specific EULA terms are not public for most of these products, which is itself a revealing data point, because when you cannot read the liability terms of the tool reviewing your building plans, the market has a structural transparency problem that no amount of procurement diligence will fix.

Municipalities will frame the AI as a tool rather than a decision-maker, arguing that the final approval stamp came from a building official whose discretionary choice to rely on algorithmic output falls within the governmental immunity doctrine just as completely as the old discretionary choice to hire a particular reviewer and trust that reviewer's judgment. This argument, frustratingly, will probably work in most jurisdictions.

Architects will point to the approval and argue that the city's reviewing authority, whether staffed by a person or a processor, found the plans compliant with applicable codes, and that municipal plan approval has historically served as evidence of code compliance in E&O disputes, a precedent that doesn't evaporate simply because the city's review process now runs through a neural network instead of a senior plan examiner with 20 years of experience and a dog-eared copy of the IRC on the desk. This argument, supported by decades of E&O case law treating municipal approval as evidence of compliance, has genuine teeth.

Homeowners get the cracked foundation and the privilege of navigating a liability framework that was not designed for this technology.

No Court Has Ruled, No Legislature Has Acted, No Insurer Has Priced This Risk

Zero courts have determined whether AI-generated plan approval constitutes a discretionary function under governmental immunity, zero state legislatures have passed statutes allocating liability for AI-assisted building code review, and zero insurance carriers have issued policy language specifically addressing AI plan review errors in residential construction. That last gap deserves particular attention because it reveals how deeply uncertain the insurance industry is about a technology they are simultaneously being asked to underwrite: AIG, Great American, and W.R. Berkley are actively seeking regulatory approval to limit their exposure to AI-related claims, citing the absence of 20 to 30 years of actuarial data on AI loss frequency and severity that would be required to price the risk rationally. Professional liability policies may not even cover AI-assisted errors under their current "professional service" definitions, and some carriers are reportedly capping AI-related claims at $500,000 on policies with $10 million face values.

Munich Re's aiSure is the only product on the market specifically insuring AI underperformance, serving all of AI, globally, with a single product. In London, Orbital Witness launched an insurance-backed guarantee for AI outputs in real estate, but its scope is limited to title and property data rather than structural code compliance. The insurance market is not ignoring AI. It is paralyzed by AI, which for a homeowner whose foundation just cracked amounts to the same thing.

Why This Gap Will Widen Before Anyone Closes It

Bureau of Labor Statistics data paints the workforce picture starkly: 147,600 building inspectors in the United States, with employment projected to decline 1% through 2034, and 14,800 annual openings that come entirely from retirements and exits because nobody new is entering the profession at a median wage of $72,120 when adjacent technical careers pay substantially more. Thirty-six percent of those inspectors work for local government, which means 36% are subject to the same budget pressures that make AI plan review attractive in the first place, creating a self-reinforcing cycle where staffing shortages drive AI adoption and AI adoption reduces the political urgency of filling those positions.

Then consider the federal layer, which evaporated in October 2025 when HUD fired all of its building inspectors: more than 90 staff members from the Real Estate Assessment Center who coordinated inspections for five million housing units in federally assisted properties, some uninspected for five-plus years because of pandemic backlogs and now uninspectable because the staff no longer exists.

This isn't a market choosing AI because AI is superior. This is a market choosing AI because the human infrastructure collapsed and the alternative to algorithmic review is no review at all, or a backlog so deep that a homeowner waits six months for a permit on a kitchen remodel. Denver didn't adopt Clariti because it was eager to automate; Denver adopted Clariti because its permitting office was drowning under a caseload it could never staff its way out of.

Florida's legislature accelerated the dissolution by passing HB 803, which waives permits entirely for single-family projects under $7,500, a threshold contractors can game by pricing jobs just below the line on paper while the actual scope creeps past it on the jobsite. Local governments lose inspection authority, insurance carriers lose the permit data they use for claims review and subrogation, and the Florida Association of Counties objected to no avail. The bill awaits the governor's signature.

In Fairness, the Machines Might Be Better Than What They Replaced

An overworked human reviewer processing their 40th plan submission of the week, fighting a backlog measured in months, with three open positions on the team that the city can't fill at $72,000 a year, is not the careful code-compliance professional that liability law imagines when it grants governmental immunity for "discretionary functions." That reviewer is scanning, checking the obvious items, and missing the Section R403.1.4 footing depth issue exactly the way an improperly calibrated algorithm would miss it, except the algorithm at least applies the same ruleset every time it runs without the degradation that comes from fatigue, distraction, or the particular kind of despair that accompanies underfunded civil service work.

AutoReview.AI was built on more than a decade of academic NLP and machine learning research at the University of Florida, designed specifically to handle the vague and flexible language that makes code review genuinely difficult, provisions like "adequate bearing capacity" that require interpretation rather than simple lookup. The system holds a US patent (No. 12,229,480, issued February 18, 2025) for this capability, and it is emphatically not a chatbot reading a PDF.

But consistency is not correctness, and even AI that is correct 98% of the time will be wrong 2% of the time, and 2% of a million annual reviews is 20,000 missed violations, each one potentially attached to a cracked foundation and a homeowner who can't find anyone in the liability chain willing to accept responsibility for the failure.

What You Should Do Before Breaking Ground

Ask your building department whether AI was involved in reviewing your plans. Most departments won't volunteer this information, and some staff may not know the answer, but if they use Clariti, AutoReview.AI, Symbium, CivCheck, CodeComply.Ai, or Archistar, the answer is yes and you should proceed with that knowledge shaping every subsequent decision in the project.

Ask your architect whether their errors-and-omissions policy covers reliance on AI-assisted plan review. Many policies haven't been updated since AI entered the permitting workflow, and if the policy defines "professional services" narrowly enough to exclude reliance on algorithmic code compliance verification, you could discover at the worst possible moment that the coverage you assumed existed has a gap precisely where you need it most. Get this in writing before you sign the design contract.

Budget $2,000 to $5,000 for an independent plan review by a licensed engineer. On a $500,000 residential build, that's 0.4% to 1% of total project cost, and if the independent reviewer catches a footing depth violation that the AI missed and that would have led to an $80,000 foundation repair, the math is so straightforward that arguing against it requires a kind of optimism that construction projects rarely reward.

Talk to your homeowner's insurance agent specifically about AI plan review scenarios. Standard policies cover structural defects discovered after purchase, but coverage disputes involving code violations that were "approved" by the permitting authority are untested territory when the approving entity was an algorithm, and you want your carrier thinking about this before you need to file the claim rather than after.

Limitations of This Analysis

No court has decided any of the liability questions raised here, making this analysis entirely prospective, built on existing tort doctrine, governmental immunity case law, and standard insurance policy language applied to a technology that has not been litigated. Vendor EULA terms are not publicly available for most AI plan review products, so the vendor liability discussion relies on standard SaaS contract patterns rather than verified specific terms. Governmental immunity varies dramatically by state, and some states have waived immunity more broadly than others in ways that could alter the analysis substantially. Most critically, no published data exists comparing AI plan review error rates to human reviewer error rates, and it is entirely possible that AI catches more violations than the average human reviewer. But "better than the average human" and "liable when wrong" are questions in different legal universes, and only the second one matters when your foundation is cracking and your contractor is pointing at the city, the city is pointing at the vendor, the vendor is pointing at the EULA, and you are standing in your living room wondering who you can actually sue.