In Honolulu, a residential building permit used to sit in a queue for months. The city’s permitting department brought in an AI tool to prescreen applications, and reviewers cut the time spent on each residential plan from 60 to 90 minutes down to 15 to 20, a drop of about 70%, according to figures the department’s director reported to GovTech. A prescreening backlog that once meant a six-month wait now clears in roughly a week.

Honolulu is not alone. Over the past year, Los Angeles, Austin and Seattle have all moved AI into the early stages of building-plan review, and the speed gains are real. So is the problem they address: an aging, shrinking municipal workforce buried under a rising volume of permits. The harder question arrives once the software starts clearing plans. When an AI prescreen waves through a design that later turns out to break code, who is on the hook?

The answer is less alarming than it first sounds, and more unsettled than the cities adopting these tools have publicly worked out.

Where AI plan review actually stands

The clearest case is Honolulu. The city’s Department of Planning and Permitting launched a tool from a company called CivCheck in December, after a five-month pilot. It checks residential applications for missing documents and code conflicts before a human reviewer ever opens the file. The point, the department’s director said, is to stop incomplete plans from triggering the repeated review cycles that clogged the pipeline in the first place.

Austin adopted a different vendor, Australia-based Archistar, in October 2024 after a three-month pilot, and now runs it as a pre-check layer for single-family permits. Los Angeles and the surrounding county launched the same tool on April 30 last year, part of the push to speed rebuilding after the wildfires. In Seattle, Mayor Bruce Harrell signed Executive Order 2025-05 in June 2025, creating a Permitting and Customer Trust team. The order itself describes an AI pilot meant to help applicants and staff catch common errors, and it is careful to say the tool is “designed to aide though not supplant human acumen” — the city’s own words, and a phrase worth holding onto.

The reason for the rush is not hard to find. The Seattle order points to a 2023 City Auditor report that found deficiencies in the city’s permitting, including a survey in which most applicants could not identify the right person to ask for help. Delay also costs real money. A study published by Hawaii’s state economic department, drawing on Honolulu permitting data, estimated that delays on private-sector projects cost $124 million in 2022 and $77 million in 2023, around 7% of the combined value of the projects held up. The labor side is just as stark. In Honolulu, the city auditor found that an average of 2,513 permit applications rolled over unfinished every year, with reviewers who were lower paid and less trained than peers elsewhere, and a department hampered by turnover and retirements.

A permit was never a guarantee

This is the part that cools the panic. A building permit has never been a promise that a structure is sound or fully code-compliant. Courts have said so for decades. A permit is permission to build, not a warranty of the work.

On top of that, cities are heavily shielded from lawsuits over plan approval. The model building code that most jurisdictions adopt contains a provision, often numbered R104.8, that relieves a building official acting in good faith from personal liability for damage tied to an act or omission in the line of duty. State law frequently goes further. California’s Government Code section 818.6 says flatly that a public entity is not liable for an injury caused by an inadequate or negligent inspection of private property, and the neighboring section 818.4 extends similar protection to the issuance of permits. The state Supreme Court applied that immunity in Morris v. County of Marin, and later courts have read it broadly. Homeowners who assume the city stands behind every permit it issues have learned, in court, that it does not.

Swapping a human prescreen for an AI one does not, by itself, dissolve that immunity. A city that uses software to flag problems faster is still doing the same legal act it always did: reviewing a plan and issuing, or withholding, a permit. On that narrow question, the law has not moved.

Where the risk actually moves

The exposure does not disappear. It shifts to places the old immunity rules were not written for.

Start with the vendor. The check is now performed, in part, by a private company’s software. The companies running these tools are emphatic that the machine does not have the final say. CivCheck’s chief executive told GovTech that “our AI doesn’t make decisions” and that the platform is built to augment reviewers, not replace them. That human-in-the-loop design is doing real legal work, because it keeps a city employee in the decision and preserves the immunity that attaches to official judgment. The moment a city lets an automated prescreen stand in for that judgment rather than inform it, the legal ground gets softer, and no court has mapped it yet.

Then there is the contract. Across government, the terms that decide who eats the cost of an AI error are being rewritten right now. In early 2026, the federal government’s purchasing arm posted a proposed clause, GSAR 552.239-7001, setting terms for AI systems used in or supplied under federal contracts, covering data rights, performance standards and the reach of a contractor’s responsibility for the AI it relies on. It is the first stand-alone attempt of its kind, and the fact that it is still in draft tells you how unsettled the ground rules are. Cities buying permitting software face the same questions the federal clause is trying to answer: who indemnifies whom when the tool is wrong, who may audit how it decides, and what the vendor actually warranted. Whether a given city secured any of that is a question that can only be answered by reading its contract.

There is also a fairness angle worth watching. By one industry estimate, roughly 60% of a standard residential plan review is objective code-checking that software handles well, while the remaining 40% calls for human judgment. If an automated prescreen ever treated applications from different neighborhoods unevenly, the result could look a great deal like discrimination, and that is the sort of claim that runs through the courts and fair-housing law, not through whatever a procurement contract happens to say.

Standout fact: In a Honolulu pilot, AI prescreening cut residential plan review from 60 to 90 minutes per application to 15 to 20, the city reported, yet the software still does not approve or reject a single permit.

What cities signed without debating

This is the quiet part. City councils approved these tools as efficiency measures, and the public conversation centered on speed: shorter queues, faster housing, relief for overworked staff. The allocation of risk, who pays when the software is wrong, lives in procurement contracts that were rarely the subject of any council debate.

That is not a scandal. It is a gap. The indemnification terms, the audit rights, the limits on a vendor’s liability, the question of what happens to the records of an AI prescreen if a dispute lands in court, are all decided in documents that a resident would have to file a records request to see. The cities moving fastest are, in many cases, the ones with the worst backlogs and the most pressure to act. Speed was the selling point. The fine print came along for the ride.

Records retention deserves its own mention, because it is the kind of detail that only matters after something goes wrong. If a building fails an inspection two years on and the owner argues the city’s prescreen missed an obvious code violation, the case may turn on what the AI flagged, what it did not, and whether anyone kept that output. The federal AI clause now in draft devotes real attention to how government data tied to an AI system is handled and deleted, which is a sign that retention is a live question wherever these tools are bought. A permitting system that quietly discards its own reasoning leaves a homeowner, and the city, with nothing to examine, and few procurement contracts say a word about it.

The Bottom Line

If you are a city official, read the indemnification clause in your AI permitting contract before the first bad approval, not after. If you are a developer, faster AI review shortens your timeline, but it does not shift your own duty to build to code, so do not treat a quick approval as cover. And if you own the home, a permit cleared in days is still what a permit always was: permission to build, not a promise that everything behind the walls is right.

Source link