An AI Checked Your Building Plans Against the Zoning Code in Three Minutes. The City Reviewer Who Checks the Other 2,000 Rules Still Needs Eight Weeks.
AI permit pre-screening tools from Archistar, CivCheck, and Blitz AI can verify zoning setbacks, lot coverage, and document completeness in minutes. Seattle tested one on real residential applications and measured 92% compliance accuracy. But accuracy and coverage are different problems. Catching 92% of the errors does not eliminate a single review cycle if the remaining 8% still requires a human to send the application back. Across 741 U.S. cities, average initial permit review takes 22.9 days. The AI works. The regulatory complexity underneath it is the actual bottleneck.
A homebuilder in Virginia submitted plans for a single-family residence to the county building department. Midway through review, the assigned reviewer resigned. Nobody handed the file to a successor. His permit sat in a dead queue until someone noticed.
That is not a technology problem. It is not even a staffing problem, exactly. It is a systems problem, and it is the kind of failure that AI permit tools are supposed to prevent. They do prevent some of it. What they cannot do yet is prevent enough of it to change the timeline that matters to the person paying the mortgage on an empty lot.
Three Minutes and 172 Pages
Archistar, an Australian proptech company, demonstrated its eCheck platform on a 172-page construction document set. Computer vision parsed the architectural drawings. Machine learning algorithms mapped every element against the municipality's zoning and building codes: setbacks, height limits, lot coverage, window placement, energy specifications. The platform flagged a missing boundary setback dimension that would have required two days of manual checking by a trained draftsman.
Three minutes. Done.
Archistar operates in more than 25 municipalities worldwide, including Vancouver, Austin, and Los Angeles. In October 2024, Austin signed a five-year contract with the company after a three-month pilot that tested eCheck on single-family residential plans. Jose Roig, director of Austin's Development Services Department, told Cities Today that the software "accurately understand[s] zoning code and implement[s] those regulations when reviewing a set of residential building plans." Results were "reliable and repeatable across a wide range of residential projects, regardless of their zoning."
Governor Newsom announced in April 2025 that Archistar would provide its pre-check tool free of charge to LA-area jurisdictions rebuilding after the Eaton and Palisades fires. The International Code Council, the organization that publishes the model codes used by most U.S. jurisdictions, formalized a partnership with Archistar to standardize automated code compliance review.
These are real deployments backed by real results. They are also, by the companies' own admission, incomplete.
The 80 Percent Starting Line
Ben Coorey, Archistar's founder, described the platform's accuracy trajectory in a company blog post with unusual candor for a technology vendor. A new completeness check module, deployed for the first time in a municipality Archistar has never worked with before, starts at roughly 80% accuracy. The company then works with the local building department to close the remaining 20% gap. With supervised learning and municipality-specific training, accuracy reaches 90% and beyond.
"We don't profess to say that our AI solution is perfectly accurate out of the box, which is why the human in the loop is still critical in what we're doing," Coorey wrote.
That is an honest assessment. It is also worth taking seriously as a constraint. An AI system that starts at 80% accuracy and improves to 90% with training is useful. On a building plan with 50 checkable items, 90% accuracy means 5 missed items. On a plan with 200 checkable items, it means 20. A single missed item that triggers a correction letter from the city restarts the clock on that review cycle.
Seattle Ran the Experiment
Between March and October 2025, Seattle's Innovation and Performance team partnered with the Department of Construction and Inspections to conduct a proof-of-value test of CivCheck, an AI pre-screening platform developed by Clariti. They tested it on real applications for small residential projects: townhomes, single-family homes, and accessory dwelling units. CivCheck analyzed the submissions. City staff reviewed the same applications independently. Both teams compared results.
They also interviewed 20 former permit applicants and more than a dozen city staff members.
Application completeness checks were 87% accurate. Design compliance checks were 92% accurate. Both numbers improved over the course of the test as CivCheck calibrated to Seattle's specific requirements.
But the report, published in June 2026 on the city's Building Connections blog, drew a distinction that most coverage of AI permitting tools has ignored. Accuracy measures whether the AI gets the right answer on the checks it runs. Coverage measures whether the AI runs enough checks to catch everything a human reviewer would flag.
"Removing a review cycle requires the pre-screening to flag every single correction up front. If the pre-screening only catches some of the corrections needed, then the applicant may end up going through the same number of review cycles as they would without pre-screening."
That sentence is the entire story of AI permit review in 2026.
Completeness vs. Compliance
Seattle's team found that completeness checks are more automatable than compliance checks. Completeness is straightforward: does the application include a site plan? Is the structural calculation sheet present? Are the required forms filled out? These requirements relate to city processes and document standards. They can be codified into discrete, binary rules. Present or absent. Filled or blank.
Compliance is different. Compliance means comparing the proposed design against the full scope of Seattle's building codes. Hundreds of rules govern setbacks, height limits, lot coverage, and floor area ratios. Hundreds more address structural engineering, energy performance, fire separation, accessibility, and stormwater management. Some of those rules interact: a setback exception on one side of the lot may trigger a different fire separation requirement. Some rules require professional judgment: "adequate natural light" is code language, but what counts as adequate depends on room orientation, window area calculations, and neighboring structure heights.
The report concluded that "some of those requirements are conceptually complex and likely difficult to automate." It added that "simplifying the City's own codes and processes would make it easier to automate both completeness and compliance checks in the future."
Read that again. The city's own innovation team said the building codes are too complex to fully automate, and suggested that fixing the codes would be more effective than improving the AI.
22.9 Days Is the Average
PermitPlace, a commercial permit expediting firm, published in March 2026 the most comprehensive analysis of U.S. building permit review timelines ever assembled. They analyzed published department guidelines from 741 cities across 44 states.
National average initial permit review time: 22.9 days. Median: 14 days. The gap between mean and median reveals a heavy right skew, with a small number of extremely slow jurisdictions pulling the average up.
Chicago is the slowest major metro at 92 days for initial review. San Francisco takes 60 days. Portland, 51. New York, 30. At the other end, Denver, Houston, and Miami all list 2-day initial intake timelines.
Eighteen cities list 180-day timelines, most of them statutory maximums in smaller jurisdictions. PermitPlace noted that actual review in these cities is often faster than published guidelines, but without active management, projects frequently hit the upper range.
Two critical caveats. First, these are published timelines for initial review only. They do not include corrections cycles. PermitPlace reported from 20 years of experience that "the gap between published timelines and actual commercial project timelines can be 2-5x." A city that lists 14-day initial review may take 28 to 70 days from submission to permit in hand. Second, each corrections cycle adds its own wait time. In Aspen, Colorado, a major new residential permit requires 16 to 18 weeks for round one review alone. Each additional review round adds 8 to 10 weeks. A project with two rounds of corrections can spend 8 months in permitting.
Denver Spent $4.6 Million on Pre-Screening
In March 2026, the City and County of Denver approved a five-year, $4.6 million contract with CivCheck. It automatically flags missing documents, incomplete fields, and application errors before plans reach city reviewers, giving applicants the chance to correct problems before formal submission.
Denver had already cut processing time for single-family and duplex projects by roughly 45% since 2023. It launched a dedicated Permitting Office and set a 180-day shot clock for permit decisions, promising to refund developers up to $10,000 in application fees if the deadline was missed.
In the same budget cycle, Denver's Community Planning and Development Department cut 59 budgeted positions, bringing its total to 251.
Julia Richman, Clariti's vice president of government relations, described the logic: "Most plan review delays start upstream, when submissions enter the queue incomplete or inconsistent." CivCheck addresses the upstream problem. An application that arrives complete and consistent requires fewer back-and-forth cycles between staff and applicant.
But Denver's investment reveals the tension. AI pre-screening is being deployed alongside staffing reductions. It does not make the code simpler or the review faster for applications that are already complete.
What Pre-Screening Cannot Reach
A homeowner planning a kitchen addition submits plans to the building department. The AI pre-screener verifies that the structural calculation sheet is present, the site plan shows the correct setbacks, and the energy compliance form is filled out. All completeness checks pass.
Plans move to human review. A structural reviewer finds that the beam specification does not account for point loads from a removed bearing wall. A fire reviewer notes that the new layout places a gas range within 18 inches of a combustible wall, violating the residential code. An energy reviewer determines that the added window area on the south elevation pushes the home's total glazing above the prescriptive limit, requiring a performance path energy calculation the applicant did not include because the pre-screener did not flag it.
Three corrections. Three different departments. Back it goes to the applicant. Wait time for the next review round: 3 to 8 weeks, depending on the jurisdiction.
AI caught the easy problems. The expensive problems are the ones that require interpreting how two sections of code interact when a wall moves.
Subjectivity Compounds the Delay
An Independent Institute analysis published in March 2026 identified a problem that AI pre-screening cannot address at all: reviewer subjectivity. Different plan reviewers in the same department may interpret the same code provision differently. One reviewer may accept a window area calculation using the simplified method. Another may require the detailed method. One may approve a foundation detail as standard. Another may flag it for engineering review.
"Subjectivity compounds the issue," the analysis noted. "Different reviewers may interpret the same code differently, with some being 'tough' and others not (just like baseball umpires with strike zones)."
An AI system trained on one reviewer's interpretations will disagree with another reviewer's interpretations on edge cases. Those edge cases are exactly the ones that generate corrections cycles. The pre-screener says the plan is compliant. The assigned reviewer disagrees. The applicant receives a correction letter citing a provision the AI did not flag.
Building codes are partially discretionary by design. "Adequate," "sufficient," "appropriate," and "as determined by the building official" appear throughout the International Building Code and its residential counterpart. These words exist because code writers recognized that physical conditions vary. They also guarantee that no automated system can achieve 100% coverage.
Naples's First, Seattle's Caution
Naples, Florida, became the first city in the state to partner with Blitz AI in 2026, integrating its automated compliance platform with the CityView permitting system. Mayor Teresa Heitmann called it "a significant milestone in our commitment to innovation, efficiency, and service excellence."
Seattle's report was more measured. After seven months of testing and months of analysis, the city published findings that acknowledged both the potential and the limitations of AI pre-screening. It committed to exploring "options for funding and logistical support to leverage new tools like CivCheck" while simultaneously promising to "continue working to clarify and streamline our permitting and decision-making processes."
One city announced a partnership with a press release. The other published a detailed technical assessment and concluded that the permitting process itself needed reform. Both are valid responses to the same technology. Only one is honest about the constraints.
What This Means If You Are Filing a Permit
If your city offers an AI pre-screening tool, use it. CivCheck, Archistar eCheck, or whatever platform your jurisdiction has adopted will catch document omissions and straightforward code violations faster and more reliably than your own review. A complete application moves faster through the queue even if it still faces corrections on compliance issues the AI could not check.
Do not expect the pre-screener to eliminate review cycles. The technology catches 87 to 92% of what it checks. It does not check everything. Plan for at least one round of corrections on any project more complex than a deck or a water heater replacement. For additions, remodels, and new construction, plan for two.
If you are in a jurisdiction where the AI pre-screener gives you a clean report, ask yourself which checks it ran and which it did not. A pre-screener that verifies zoning setbacks and document completeness has not evaluated your structural engineering, your energy compliance path, or the interaction between your proposed design and the fire separation requirements triggered by your lot's proximity to the property line.
Budget time based on real timelines, not published averages. PermitPlace's data shows a 2-5x multiplier between published initial review times and actual permit-in-hand timelines. If your city lists 14 days, plan for 30 to 60. If it lists 45, plan for 90 to 150. AI pre-screening may reduce the number of corrections cycles by one, but each remaining cycle adds weeks to months.
The honest conclusion is unglamorous but useful: AI permit tools make the process marginally faster by reducing upstream errors. They do not make the process fundamentally faster because the complexity is in the codes, not the review queue. Until jurisdictions simplify the regulations themselves, the reviewer with the red pen and the eight-week turnaround remains the rate-limiting step.