The Cities That Handed Plan Review to AI
A commercial permit in Honolulu has waited a median of 393 days to clear review. That number is why the city now runs building plans through an AI system that scans for code compliance before a human opens the file.
Honolulu is not alone. Los Angeles went live with AI plan checking on April 30, 2025, letting property owners pre-check designs against zoning and building rules in the middle of the post-wildfire rebuild that followed Governor Newsom's January 12 executive order. Seattle is building toward a 2026 rollout, routing development applications through a pilot run by a Permitting and Customer Trust team.
Austin's AI Pre-Check launched September 30, 2025, returning a zoning-compliance report within one business day. Preliminary testing showed nearly a 50 percent cut in zoning-reviewer time. The same kind of check has moved into the private market too, where platforms that read a site against the local code let a team run it before they ever submit.
The headlines read like a software story. They aren't. Why this works has almost nothing to do with how clever the systems are.
It has to do with the nature of code. A large part of any development code is just numbers. A setback is a number. A height limit is a number. Lot coverage is a percentage, and a completeness check is a count of required sheets.
Comparing the number on a drawing to the number in the code never needed human judgment. It needed a careful reader who wouldn't skip a line. AI front-loads that layer and leaves the judgment layer where it has always been: with people.
The adoption wave is real and uneven
Four cities, four different starting points. Honolulu came to it with one of the longest commercial waits in the country and a permitting back end old enough to predate the modern web. Los Angeles came to it under disaster pressure, with thousands of homeowners who needed to rebuild and no appetite for a year in queue.
Seattle came to it through an executive order and a deliberate pilot. Austin came to it through a residential program that started narrow and grew.
The scopes differ as much as the motives. Austin's Pre-Check opened only to Expedited Residential Building Plan Review applicants, checking zoning-regulation compliance and nothing else. In February 2026 the city opened an expanded beta to qualifying standard residential customers, while excluding planned unit developments, neighborhood conservation combining districts, and projects with approved site plans.
Los Angeles let owners pre-check before submission. Seattle is still routing applications through a pilot rather than a public-facing tool.
The pressure behind all of this is housing. Austin's building surge has done what supply is supposed to do. Pew reported in March 2026 that the city's wave of new construction drove rents down, a rare result among large U.S. metros.
Faster review is part of how that supply reaches the ground. A year saved in queue is a year of carrying cost a project doesn't pay, and a year sooner units come online.
What unites the four is not a product. It is a decision about which part of review a machine should touch first. Every one of these cities pointed the system at the same layer: the rules a drawing either meets or doesn't, measured against a value written in the code. None of them pointed it at the discretionary questions. That isn't a coincidence, and it isn't a roadmap waiting to be finished.
Why the measurable layer was always machine-work
Read Austin's own description of what Pre-Check does. It "will read and evaluate your design based on a defined set of development requirements." The operative phrase is "defined set." A requirement that is enumerated and fixed in advance is one a machine can check, because checking it means reading a value off a drawing and comparing it to a value in the code.
This is the part of plan review that was always mechanical and only looked like expertise because a person had to do it meticulously, which takes time. Does the rear setback meet the minimum. Is the building under the height cap. Does lot coverage stay inside the percentage. Are all the required sheets present and legible.
A reviewer doing this work isn't exercising discretion. They're running a checklist. A checklist run by a careful human at a few plans a day is the same checklist run by a system at a few plans a minute, only slower and more prone to fatigue. The completeness check that opens any review is the cleanest example: it is a count, and a count has a right answer.
That is why the early results are real rather than hype. Austin's nearly 50 percent cut in zoning-reviewer time isn't the system being smart. It is the system being fast and tireless on work that was already enumerable.
The city is candid about this. Senior Public Information Specialist Stephanie Sanchez put it plainly: "We see AI as a powerful support tool that can improve both speed and customer service in our permitting process." A support tool. Not a decision-maker.
The line the systems do not cross
There is a second kind of rule in every code, and it doesn't reduce to a number. A variance turns on a hardship finding, a judgment about whether a specific property carries a specific burden that justifies relief. A cross-discipline conflict turns on which requirement gives way when two of them collide on the same site. Negotiated relief turns on what a reviewer and an applicant agree to after a conversation. None of these is enumerable in advance, because the input is the particular site and the particular argument, not a value on a sheet.
A machine doesn't check these, and the reason is structural. This isn't a feature the systems forgot to ship and will add next year. The discretionary layer is, by definition, the part of the code written to require a human to weigh facts the code couldn't anticipate.
You can't enumerate a hardship finding the way you enumerate a setback. The whole point of the finding is that it depends on circumstances the drafters left open.
The compounding constraints that pile up across disciplines are the same story. When drainage, fire access, and tree protection each pull a site plan in a different direction, resolving the conflict is an act of judgment about trade-offs, and trade-offs are exactly what a conformance check is silent about.
This is why the cities describe their systems as sitting inside the human review process rather than replacing it. The human reviewers keep final authority, and the AI feeds them a conformance report rather than a verdict.
Austin's Pre-Check sits at the front of a residential review that still runs through people. The machine clears the numbers so the humans can spend their time on the questions only humans can answer.
A clean pre-check is not an approval
A clean conformance report comes back quickly, and the speed makes it easy to misread what you're holding. That misreading is the trap.
Consider a hypothetical. A small residential infill project, single lot, submitted through a city's pre-check before formal review. The system reads the drawings. Setbacks clear. Height clears. Lot coverage clears. Every sheet the checklist requires is present. The report comes back clean, and the applicant reads it as a green light.
It isn't. The conformance pass confirmed that the drawing's numbers match the code's numbers. It said nothing about whether the proposed driveway location conflicts with a drainage requirement that lives in a different discipline's review. It said nothing about whether a recorded plat note imposes a setback stricter than the code minimum the system was checking against, or whether the project needs a variance for a condition the numbers alone do not surface. The machine did its job perfectly, and that job was narrow on purpose.
A conformance pass tells you the measurable rules are satisfied. It doesn't tell you the project is approvable, because approvability includes the discretionary, cross-discipline, and relief questions the pre-check never examined and was never built to examine. Reading a clean pre-check as an approval is the same error as reading a passed completeness check as a passed review. Both confirm the front of the process; neither confirms the end of it.
This distinction is going to matter more as the tools spread. The reporting on Austin's AI permitting rollout and the city's broader accelerated site-plan-review reforms both point the same way. Cities are putting the measurable layer up front and keeping the judgment layer in human hands, deliberately.
Where this goes
Four cities now, and the number only climbs. AI plan-check is moving from pilots to public tools, and from residential into commercial work. The measurable layer of review is on its way to instant.
When conformance takes hours instead of weeks, the slow part of getting a project approved stops being the numbers. It becomes the judgment. The variance that turns on a hardship finding, the cross-discipline conflict that surfaces only when two rules collide on the same site: that work doesn't automate, and it doesn't speed up. It becomes the whole game.
For a generation, the mechanical check and the judgment call sat tangled in the same reviewer's queue, which made them look like one job. They were always two. The cities adopting AI have pulled them apart, and they aren't going back.
So the skill that matters is shifting. A clean pre-check is becoming table stakes. Knowing what the pre-check can't see is becoming the job.