Six Cities Let an AI Review Your Building Plans. None of Them Agreed on What Happens When It's Wrong.

Before December 2025, a residential building permit application in Honolulu sat in a prescreening queue for roughly six months. DPP processed approximately 20,000 building permit applications a year and authorized about 1,400 new private housing units, which meant the bottleneck wasn't the construction. It was the paperwork before the construction, and specifically the back-and-forth between applicants who submitted incomplete plans and reviewers who sent them back with an average of 23.5 corrections apiece.

In December, Honolulu launched CivCheck, an AI-guided permit preparation tool built by Canadian company Clariti. Seven months later, the prescreening backlog is roughly one week. Applications prepared through CivCheck reach a permit decision in 32.5 days, compared to 73 days without it. Corrections per permit dropped from 23.5 to 7.7. Reviewers who once spent 60 to 90 minutes on each residential plan now spend 15 to 20.

"It's kind of like TurboTax for permitting," Dawn Takeuchi Apuna, DPP's director, told StateScoop. "The AI really helps guide applicants through a much better quality application, and then in turn it means less review cycles and faster turnaround."

Honolulu is not running this experiment alone. Over the past eighteen months, at least five other cities and counties have moved AI into building plan review, each with a different vendor, a different scope, and a different set of assumptions about what happens when the technology is wrong.

The Map So Far

Sources: Honolulu DPP, StateScoop, HousingWire, Clariti press release May 2026, Seattle Executive Order 2025-05. Scope and status reflect most recent public information as of July 2026.

In Florida, three additional municipalities are using AutoReview.AI, a University of Florida spinoff that automates compliance checking against the entire Florida Building Code. Pasco County, Gainesville, and Altamonte Springs have each deployed it, with Altamonte Springs City Manager Frank Martz noting that "AI can do it in as little as 30 minutes" what previously took a staff member several weeks. Behind AutoReview.AI is more than a decade of academic research, including a patent-pending natural language processing framework designed to interpret the ambiguous sections of building codes that have historically made automated checking so difficult.

Speed Is Real. Legal Frameworks Are Not.

Nobody disputes the numbers. Honolulu's 67% reduction in corrections per permit and 55% faster time-to-decision are official figures from DPP, published on the department's own website and confirmed by its director in multiple interviews. On efficiency, the case is settled.

What nobody has settled is what happens when the efficiency case collides with a code violation that the AI prescreened and missed.

Here is the current legal arrangement, stripped to its mechanics: a building official personally certifies that a permitted structure complies with applicable codes. That certification creates legal liability. If the building later fails and someone gets hurt, the certification is the document that lawyers examine. A building official signed it. An architect or engineer of record stamped the plans. Those two signatures are where responsibility lives under existing law.

AI plan review, as currently deployed, sits outside this framework entirely. Every city using it has been careful to say so. Honolulu's CivCheck page states explicitly that it "does not replace the DPP's official review, nor does it guarantee permit approval." Seattle Mayor Harrell's Executive Order 2025-05 describes the AI pilot as "designed to aide though not supplant human acumen." CivCheck CEO Dheekshita Kumar has repeatedly emphasized the tool's "human-in-the-loop" design, calling it a system that augments plan reviewers rather than replacing them.

These disclaimers are accurate descriptions of how the tools currently work. They are not answers to the question of what happens in three years when the tools have been working well enough that cities begin treating them as de facto reviewers rather than prescreening assistants, which is the trajectory every adoption curve in this table is pointed toward.

The Honolulu Precedent

Honolulu is the furthest along and therefore the most instructive case. DPP launched a "Priority Review" program in July 2026 that routes CivCheck applications ahead of non-CivCheck applications in the review queue. DPP has announced that CivCheck will become mandatory for residential permits later this year. Once that happens, every residential application in the county will pass through Clariti's AI before it reaches a human reviewer, and the practical question becomes how closely the human reviewer examines a plan that the AI has already flagged as substantially compliant.

Behavioral research on automation bias suggests the answer is: not as closely as they should. When a system consistently produces correct results, human operators develop justified confidence in it, and that confidence reduces their scrutiny of the system's outputs. CivCheck's entire purpose is to produce correct results more consistently than unassisted applicants, and the 67% reduction in corrections proves it is doing exactly that. Better AI performance breeds less careful human checking. Not a failure of the humans; a predictable consequence of good automation.

Liability sharpens here: if a reviewer spent 15 minutes on a plan because CivCheck flagged nothing, and the plan contained a fire egress violation that CivCheck's model wasn't trained to catch, does the city's liability exposure change because the reviewer relied on an AI prescreening tool? Under current law, probably not. Yes, the building official still signed the certification. But under current practice, the reviewer's behavior was shaped by a tool the city mandated, and the tool's training data and accuracy rates are proprietary to a Canadian software company whose contractual obligations to the City of Honolulu have not been publicly disclosed.

The Counterargument Deserves Full Weight

There is a strong case that AI plan review reduces total liability exposure rather than increasing it. Pre-CivCheck, Honolulu applicants submitted plans with an average of 23.5 errors. Post-CivCheck, that number is 7.7. Some of those errors were minor labeling issues, but others were substantive code violations that would have reached the construction phase if missed by an overworked reviewer processing the 50th plan of the month after three people on their team retired and weren't replaced, which is a description that fits most municipal building departments in the United States right now.

AI catches more errors, earlier, than the understaffed status quo. If you are quantifying risk rather than assigning blame, the AI-assisted pipeline is almost certainly safer than the pipeline it replaced. A 2019 survey of Florida building professionals identified the most frequently violated code sections as means of egress, fire resistance, glazing, and roof ventilation. These are precisely the kinds of pattern-matching compliance checks that machine learning handles well, and they are precisely the violations that get people killed when they go undetected.

Where this argument breaks down is scale. When a human reviewer misses a code violation, one building is affected. When an AI model has a systematic blind spot, every building that passes through that model during the period of the blind spot is affected simultaneously, and the defect may not surface until something fails. Risk shifts in kind, not just in frequency, and existing legal frameworks were designed for the human version.

What California Is Watching

California Assembly Bill AB-6 requires the Department of Housing and Community Development to convene a working group by December 31, 2026, to consider amendments that would allow residential developments of 3 to 10 units to be built under the California Residential Code. AB-6 also requires the department to review construction cost pressures from building standards and report findings to the Legislature. While AB-6 does not mention AI plan review directly, the working group's mandate to reduce construction costs and streamline permitting creates the regulatory surface on which AI tools would logically be deployed statewide, and California's decisions tend to propagate.

Meanwhile, in the academic pipeline, researchers at the University of Florida published an AI system called LLM-QueryBC in the ASCE Journal of Construction Engineering and Management that uses large language models to query building codes in natural language, achieving 24% better accuracy on textual regulations and 25% better performance on complex table-based queries than existing methods, with response times under 10 seconds. Researchers validated it against a 15-story mixed-use high-rise. If you want to see where the next generation of these tools is headed, the academic literature is moving from "can AI check codes?" to "how accurately can AI interpret ambiguous code provisions?" and the answer is getting better faster than the legal frameworks are adapting.

What This Means If You're Filing a Permit

If you're a residential builder or homeowner filing a permit in Honolulu, Austin, LA, Seattle, Denver, or Naples, check whether your jurisdiction offers AI-assisted prescreening and use it. In Honolulu, CivCheck is free, currently optional, and will be mandatory soon. Using it correlates with a 55% faster permit decision and dramatically fewer correction cycles. Practically, the benefit is substantial and immediate.

If you're an architect or engineer of record, the AI pre-check does not change your professional liability. Your stamp is still your stamp. But it may change your workflow: if the city's AI is catching code issues before your plans reach formal review, you should be running the same tool yourself before submission rather than discovering its findings in a correction letter. Several jurisdictions, including Honolulu, make the tool available to design professionals specifically for this purpose.

If you're a building official, the harder question is yours: when the AI consistently catches things you would have missed, and when staffing levels make it physically impossible to review every plan with the same rigor you applied in 2019, what does "professional certification" actually mean? Nobody building these tools will answer that for you. Nobody using these tools has been forced to answer it yet. When the first building fails in a jurisdiction with mandatory AI prescreening, someone will have to.

What This Doesn't Resolve

This analysis does not address commercial plan review, which involves significantly more complex structural and MEP coordination and where AI tools are earlier in deployment. Honolulu expects to extend CivCheck to commercial projects by mid-2026, but no performance data exists for that use case yet. Efficiency gains reported here are residential-specific and should not be extrapolated to commercial without independent verification.

Our liability analysis relies on publicly available information about the contractual terms between cities and AI vendors. Actual contracts may contain indemnification clauses, limitation-of-liability provisions, and insurance requirements that materially affect the risk allocation. None of the six cities analyzed here have published these contracts in response to public records requests, and the question of whether they should is itself a policy question that has not been asked in any city council meeting we could find a record of.

Most importantly, no jurisdiction covered in this article has reported a case where AI prescreening cleared a plan that later resulted in a code-violating structure. Every liability concern described here is theoretical. It will not stay theoretical forever, but as of July 2026, the safety case for AI plan review rests on an absence of reported failures rather than a proven ability to prevent them at scale over time. That distinction matters, and honest reporting requires saying so.

Catherine Chen covers policy and regulation for AI Home Building. She has no financial relationship with Clariti, CivCheck, Blitz AI, AutoReview.AI, Archistar, or any vendor mentioned in this article.