Your City Now Lets AI Read Your Permit Application First. Run It Before the City Does.

In Honolulu, the average building permit used to take 73 days. Then the city started letting applicants run their plans through an AI prescreen called CivCheck before formal submittal, and the average dropped to 32.5 days. Average review cycles fell from 3.4 to 1.4. The Department of Planning and Permitting says the prescreen backlog that stood at five weeks in December is now seven days. Those numbers come from the department itself, based on 19 residential permits against 17 comparable projects. Hold that sample size in your head while you read the rest of this.

Five jurisdictions are now in the CivCheck lane: Honolulu, Seattle, Denver, Toronto, and San Jose in a pilot. The playbook is identical in every city. Before your application enters the city's official queue, you upload your plans and documents to the AI, which checks them against the city's requirements and hands you a list of missing paperwork, incomplete fields, and likely code violations, each cited to the requirement it violates. You fix them. Then you submit. Fewer correction cycles is the promise, and each correction cycle costs you about a month. The code does not care about your feelings, but it apparently cares about your lead times.

Seattle did the most honest version of this experiment, which is why its findings deserve more attention than its press coverage. Starting in March 2025, the city ran a proof-of-value study with the Department of Construction and Inspections, comparing CivCheck's checks against the same applications reviewed by actual staff, and interviewing 20 former applicants plus more than a dozen staff members. Seattle published the results in June. Application completeness checks came in at 87 percent accurate. Design compliance checks hit 92 percent. Accuracy improved over the test as the vendor calibrated the tool.

Then Seattle said the thing vendors never say, which is the most important sentence in this article. Coverage matters as much as accuracy. Applicants save the most time when prescreening eliminates an entire review cycle. Eliminating a review cycle requires flagging every single correction up front. If the prescreen catches most of the problems but misses one, the applicant goes through the same number of cycles anyway, just with a head start. Completeness checks are easy to automate because they check a small, fixed set of document requirements. Compliance checks mean comparing a design against thousands of code provisions, some of them conceptually complex, and Seattle concluded that automating away whole cycles on compliance is harder. One developer told the city he would rather wait a day or two fixing prescreen flags than burn a month in a real review cycle. Time is everything, and he is right, which is exactly why the coverage caveat matters so much.

Toronto is the newest arrival, launching its Building Permit Application Pre-Check on August 30, starting with residential projects, aimed at a provincial target of 285,000 new homes by 2031. The announcement cites a 55 percent cut in time to permit decision for Honolulu residential permits in the first quarter of 2026. Note the chain of custody on that number: a vendor announcement citing a city department citing a 19-permit sample. Directionally encouraging. Not a building code.

Denver, which selected CivCheck in March, included the clause lawyers should read twice: final permit approvals and code interpretations remain with the city's licensed professionals and subject matter experts. The AI is a screening layer, not an approval authority. Honolulu's CEO-speak version of the same point comes from CivCheck CEO Dheekshita Kumar, who says the tool is designed to augment plan reviewers, not replace them. Both statements are true and both are doing quiet work: the moment an AI's flag becomes dispositive on a code interpretation, the city has delegated its police power to a vendor, and nobody in any of these announcements wants that conversation.

San Jose illustrates the actual disease this treats. More than 90 percent of accessory dwelling unit applications there are returned to applicants for missing information. Ninety percent. The prescreen tool targets the paperwork failure, not the design failure, because the paperwork failure is what generates most of the cycles. San Jose's pilot plans to release the tool to the public first for ADU permits, which account for roughly a quarter of new-build permit applications. If nine out of ten applications are incomplete at first submission, the AI prescreen is not really a technology story. It is a checklist story with good PR.

That is the skepticism, stated plainly. The samples are small and self-reported. There is selection bias: applicants who bother to run a prescreen may submit stronger applications regardless. The tool cannot adjudicate judgment calls, variances, historic design review, or ambiguous code language, which is precisely where human reviewers earn their skepticism. Seattle's own conclusion was that the city has more work to do to make its requirements clearer and less subjective, which is a polite way of saying that some of the code is not automatable because it is not quite readable. And there is a regulatory shape to watch: Honolulu is phasing in mandatory CivCheck, which turns an optional prescreen into a toll booth. Once the checker is mandatory, applicants optimize for the checker instead of the code, and the city becomes dependent on a vendor's model for the front door of its permitting process. Cities should write that contract carefully. Applicants should not assume the mandatory version will be as friendly as the voluntary one.

Still, for the person holding a project, the practical guidance is simple and it survives the skepticism. If your city offers the prescreen, run it before you submit. It catches the missing documents and incomplete fields that cause most first-round rejections, and it does that job at 87 to 92 percent accuracy. Obsess over completeness, not cleverness. Every required form, stamp, sheet, and signature, present and legible. The tool's highest coverage is on the boring stuff, which means the boring stuff is now the highest-leverage place to spend your preparation time. Do not read an AI pass as an approval. Expect at least one human review cycle for anything touching real code interpretation. Budget on the Honolulu numbers only as directional: roughly 33 days with a prescreen versus 73 without, on a small sample, from the department promoting the tool. And if your city does not have AI prescreening at all, the underlying lesson is free: most permit delays come from incomplete submittals, not from design disputes. Ask your building department for its completeness checklist before you file. It is the oldest automation in permitting, and it still works.

What this analysis did not prove: the Honolulu figures (73 to 32.5 days, 3.4 to 1.4 cycles) come from the department's own reporting on 19 CivCheck permits versus 17 comparable projects, and the sample is too small for statistical confidence. Toronto's 55 percent figure is cited in a vendor press release from the same pilot. Seattle's accuracy figures (87 percent completeness, 92 percent compliance) are from the city's own test and improved during calibration, but the city also found that partial coverage does not reduce cycle counts. No source establishes how these tools perform on complex commercial projects or on code interpretations involving variances and historic review.