Automated Permit Checking: Pre-Screening Is Running, Full Code Compliance Is Not Shown

The problem

The anchor paper on contract review¹ reported that language models flagged departures from a written playbook at close to junior-lawyer accuracy in a ten-contract test whose authors worked for a legal-technology company (measured without replication, January 2024)², while redlining, the proposing of edits, trailed a lawyer baseline by 15 to 26 points on 20 samples (benchmark, February 2025)³. It placed the use case in early use, with thin and not fully independent evidence¹. This paper adds a different problem: the material checked is a set of drawings, forms and site data, the rules differ from place to place, and the result feeds a public decision with a deadline.

In England, councils decided 19% of major planning applications within the statutory 13 weeks in April to June 2026, although 91% were decided within 13 weeks or an agreed extension (official statistics, measured)⁴. For minor applications the figures were 39% within 8 weeks and 87% within 8 weeks or an agreed time⁴. There were 113,800 applications on hand at 1 April 2026, 53% above the number decided in the quarter⁴. These figures cover planning permission, which comes before building-permit plan checking. The release gives no figure for how many applications are returned as incomplete, and none for how much decision delay holds up homes⁴.

In the United States, a Denver city official said only about 37% of applications are accepted on the first round (city-reported, unnamed official, 2026)⁵. Honolulu's residential permitting averaged eight months as of early 2025, according to Honolulu Civil Beat reporting cited by HousingWire (media-reported)⁶. A study of Los Angeles approvals from 2010 to 2022, published in the Journal of Urban Economics in August 2026, estimated that cutting approval times by 25% could have raised housing production by almost 24% (modelled estimate, reported by Stateline; the study itself was not read)⁵.

What already works

In Honolulu, the 73-to-32.5-day figure comes from city and vendor officials, and HousingWire linked it to a vendor page⁵⁶. A separate Civil Beat review of city data found the median wait for a new building permit fell to 2.5 months from August to April, a 40% decrease on the year before, but the city's overhaul also replaced its permitting platform, so the figure does not isolate the AI check⁵. Review of this kind currently covers smaller residential projects only⁶. In every deployment found, human reviewers keep the final decision⁵⁶⁷.

Seattle's accuracy figures are the only error measures found. The report's unit of counting was not established, so it is unclear whether 87% and 92% describe individual checks or whole applications. The reported gains are real enough to be worth testing, but they are not yet independently confirmed.

In Singapore, submissions through CORENET X became mandatory on 1 October 2025 for new projects of 30,000 m² gross floor area or more, and from 1 October 2026 for new projects of 5,000 m² or more; smaller projects may stay on the older system after industry feedback (official circular, 23 July 2026)⁸. Building models must use an open format called IFC+SG⁸. The Model Checker validates these models against basic standards, modelling standards and "selected regulatory requirements", with examples such as insufficient headroom and undersized bicycle parking, and it is being introduced in phases, starting with quality checks⁹. The operators report that gains are most pronounced for larger and more complex projects⁸. No page read gave accuracy figures for the Model Checker or said that it uses machine learning⁹.

A narrower precedent exists in the United States. SolarAPP+, a tool from a national laboratory for residential solar, can issue a building permit automatically, and a laboratory study of five test jurisdictions found average permit times fell to under a day (December 2021 data, older than three years, reported by a newsletter)¹⁰. The same newsletter, as of 2022, recorded earlier code-checking projects that lost funding (SMARTcodes), made little progress (AutoCodes), were no longer developed (UpCodes AI) or "didn't succeed" (an earlier Singapore effort), and attributed the narrow successes to short checklists and acceptance by stakeholders¹⁰.

What does not exist, in the sources opened: no independent evaluation of an AI permit check with a published method, no accuracy measure for automatic compliance decisions on full building designs, and no deployment that lets AI approve anything beyond narrow, tightly defined cases.

The possibility

What it could look like. If the conditions below held, an architect or builder could submit drawings, forms and site data to a pre-check before applying formally. The tool could read them, list missing documents, and flag places where the plans appear to conflict with a code requirement, citing the clause. The applicant would correct and resubmit until the list was clear. Only then would the application reach a reviewer, who would see a shorter, cleaner file and keep the decision. This is close to what Honolulu and Seattle describe⁵. The further step would be to make the code-conflict part deeper and consistent, so that a pass means the same thing for every applicant.

A narrower step further. For a small class of simple, tightly defined projects, a clean pass could trigger automatic issue of the permit, as SolarAPP+ does for residential solar¹⁰. Whether any building type beyond that is simple enough is untested in the sources found.

The role of AI. Machine learning could do three jobs here: turning scanned or unstructured drawings and PDFs into data, matching a project to the relevant clauses in long local codes written in prose, and spotting inconsistencies between documents. Much else would be ordinary systems: the submission portal, fees, workflow, deadlines, staffing, and fixed rules applied to structured building models. If building models arrived as structured data, as Singapore requires for larger projects, geometric checks such as headroom could be done with fixed rules rather than AI⁸⁹. AI would matter most where applications arrive as unstructured files. The largest gain may also come from the plain step of telling applicants early what is missing, and the evidence does not separate that step from the AI.

What would have to be true

1. Rules would have to be read consistently. A study cited in a 2022 newsletter found the same building had no egress violations in one jurisdiction and 16 in another (study date not shown, older than three years, secondhand)¹⁰. A tool tuned to one place's reading could be wrong in the next.
2. Inputs would have to be readable. Singapore requires structured models for new projects of 5,000 m² or more and left smaller ones on the older system after feedback⁸. Where applications are unstructured, an AI reading step would have to be accurate enough to trust.
3. Accuracy would have to be measured independently, by type of check. Seattle reported 87% and 92% for two kinds of check⁵. A reviewer would need to know the error rate per application, not only per check, and for which kinds of project.
4. The output would need a defined status. Every deployment read keeps final decisions with human reviewers⁵⁶⁷. The tool's result would have to be defined as advice, with clarity on who is responsible if an approved plan later proves non-compliant.
5. Someone would have to pay and staff it. Denver approved about $4.6 million over five years⁵. HousingWire reported that Honolulu invested $7 million, a figure it did not source⁶. A federal fund of $200 million a year and HUD grants of up to $3 million were reported as available in August 2026⁵. Whether smaller authorities could afford it is not established.
6. Delay would have to sit where the tool acts. A Honolulu official said applicants account for at least 50% of elapsed time (unsourced claim)⁵. If that held elsewhere, pre-checks would act on a large share of delay. If delay were mostly reviewers' queues, they would help less.

What could stop it

What the evidence shows

• No independent evaluation. No independent evaluation of an AI permit check, with a published method, was found in the sources opened. The Honolulu and Seattle figures are city- or vendor-reported⁵⁶, and the Florida League document gives no figures⁷.
• Cities holding back. Seattle's June 2026 report said the city was not ready to adopt CivCheck fully and was still exploring funding⁵. Louisville's pilot had not launched publicly⁵. The one dissenting Denver council member raised concern about inaccuracy, and said local architects and developers doubted AI could handle the complexity of zoning codes⁵.
• Stated limits. A Honolulu official said AI does not necessarily catch or decide the grayer areas of a project⁵. Results so far come from smaller residential projects⁶.
• Rollout friction. Honolulu's wider overhaul was "not entirely smooth"; an anonymous survey of more than 150 employees, about half the staff, was almost universally negative, as reported by Civil Beat, although it concerned the new platform rather than CivCheck alone⁵.
• Mixed signals on size. Singapore's operators reported larger gains on larger projects⁸, while Honolulu's reported results come from small ones⁶.
• History. As of 2022, earlier code-checking efforts had stalled or failed, and a 2009 US courthouse modelling project was held up by hundreds of hours spent labelling spaces (older than three years)¹⁰.
• Non-AI levers. In England, 93% of major decisions made under a performance agreement (an agreed timetable) were on time against 84% of those without one, in April to June 2026; the comparison is descriptive, not causal⁴. A recently passed Georgia bill set a 45-day deadline for initial permit reviews (reported July 2026)⁶.
• Data. A Syracuse official asked who would own and handle the data a vendor tool processes⁵.

Reasoning

If an applicant relied on a clean pre-check and a reviewer later found a violation, the applicant could bear the cost of redesign and delay, while the vendor and the city bore little. An advisory-only rule would protect the city but could cut the tool's value to applicants. If reviewers came to rely on the tool, an error rate of a few per cent could pass unseen into approvals for buildings people live in, and the risk would fall on future occupants.

Incentives could pull against the idea. A vendor paid per contract or per application would gain from wider use whether or not outcomes improved. A city that measured only speed could reward a tool for passing applications.

Simpler measures could also do part of the work: completeness checklists, pre-application meetings and fixed timetables. The reported fall in review cycles could come partly from forcing complete submissions, and none of the sources separated the AI from that. On current evidence, one reading is that AI could be a useful part of a pre-check but not yet the decisive part.

Signs to watch

• Seattle's decision on adoption and funding after its June 2026 report⁵.
• Public launch and first results of Louisville's pilot⁵.
• Denver's first-round acceptance rate, measured against the 37% baseline and the 80% goal⁵.
• The CORENET X Model Checker expanding beyond its quality-check first version, and any published results from the 1 October 2026 threshold of 5,000 m²⁸⁹.
• A published independent evaluation of Honolulu's CivCheck, with its method.
• Award announcements under the HUD grants of up to $3 million for automated permitting⁵.
• The next English planning statistics, due December 2026⁴.