Contents6 sections
Federal Funding for Automated Permitting
The U.S. Department of Housing and Urban Development is offering up to $3 million for local governments to deploy automated building code permitting systems. The notice is specific about what the money buys: a demonstration that "will test the real-world deployment of an automated permitting platform within a government permitting environment to evaluate its effects on processing timelines, workflow efficiency, staff roles, and applicant experience," producing "empirical evidence on operational performance, costs, governance needs, and potential cost savings to inform broader adoption by state and local jurisdictions."
Housing policy is driving that spending. Permitting delay has been named as a constraint on housing supply, and automation is now funded as a response. The National Association of Home Builders, an industry association, told Congress that regulatory costs at the federal, state, and local levels account for 24 percent of the final price of a new single-family home built for sale. Building departments are where that cost is either incurred or avoided.
The question we hear from directors and municipal CIOs is not whether AI pre-check works somewhere. It is what has to be true inside their own department before it works there.
What the Early Adopters Report
Baltimore, Denver, and Los Angeles have launched AI tools to streamline housing permitting, with smaller jurisdictions including Everett, Washington; Lebanon, New Hampshire; and Naples, Florida following. The tools read a submittal against local requirements and flag missing or incomplete information before it reaches a reviewer.
Denver states the problem most clearly. A city official said about 37 percent of applications are accepted on the first round, and the goal is to raise that to 80 percent with the new tool. Two out of three submittals come back before anyone evaluates them on the merits, usually over a missing sheet or an unsigned form rather than a question of code interpretation.
Honolulu has published the most operating detail. Its Department of Planning and Permitting reports average review times falling from 73 days to 32.5 days and average plan review cycles dropping from 3.4 to 1.4. DPP is candid about the base. Spokesperson Davis Pitner told the Honolulu Star-Advertiser, "We recognize that 19 completed applications is a relatively small sample," and said the department will continue to validate results using a much larger dataset. Staff reaction there has been reported as largely negative, which belongs in the evaluation alongside the cycle counts. Reviewers who do not trust a flagged-item list will re-check the full submittal themselves, and the saved cycle never materializes.
Seattle measured 87 percent accuracy on completeness checks and 92 percent on design-compliance checks during its pilot, and as of the city's June 2026 report was still exploring funding and logistical support rather than adopting the tool. All of these are city and vendor reported figures, published by parties with a stake in the result and not independently verified. We read them as early signals, the kind of evidence HUD's demonstration is designed to test rather than replace.
The Case for Completeness Checks
Completeness checking is the right first application of AI in permitting because the failure it addresses is structural. A checklist condition either is or is not satisfied. No interpretation is delegated, no approval authority moves, and a wrong flag costs an applicant a clarification rather than an improper issuance.
The volume of rework shows up in routine municipal reporting. Fairfax, Virginia publishes its permit metrics, and for new commercial permits records an average of 2.80 submissions before approval and 2.33 incomplete submissions against a total time to issuance of 26.89 weeks. Every additional round consumes reviewer hours that produce no decision.
The Foundation Underneath the Model
A pre-check tool needs something to check. Digital intake rather than PDFs arriving by email, submittal data that is structured and machine readable, and a workflow defining who sees a flag, who can override it, and how the override is recorded. Where intake is unstructured, the output looks plausible and reviewers learn to ignore it.
We described this pattern in The Last Twenty Percent Problem in GenAI Solutions. In that piece we argued generative AI reaches a convincing prototype quickly, then stalls on validation, auditability, and cases that resist clean pass-or-fail conditions. Permitting has that shape. In our view, variances, nonstandard occupancies, and conflicting code sections will keep requiring planner judgment no matter how good the tooling gets.
It is a pattern we recognize from our own work. Spruce builds licensing, permitting, and inspections systems for public sector agencies, and the modernization questions surfacing around AI in building departments (structured intake, clean data, auditable workflows) are ones we encounter regularly. The moment feels like a good one for agencies to take stock of where their systems stand, because the foundation that makes AI pre-check viable is the same foundation that makes permitting faster and more auditable with or without a model on top. Our public account of one large municipal building safety agency, overseeing zoning, code enforcement, and new construction across more than 1.1 million existing buildings, describes replacing a legacy mainframe with a platform-based solution whose permits module processes hundreds of thousands of permits annually, with processing time moving from weeks to days. To put that volume in perspective, a department processing hundreds of thousands of permits a year is handling roughly what a staff of plan reviewers working strictly by hand would need months to clear in a single backlog push. That is our own reporting rather than an independent evaluation, and it involved no AI.
Baseline Before the Award Letters
Departments should be recording three numbers now: first-round acceptance rate, correction rounds, and reviewer hours, segmented by permit type the way Fairfax reports them. First-round acceptance tells you how much of the queue is rework rather than review, the only portion a pre-check tool can move. Correction rounds tell you how long a rejected application stays in circulation. Reviewer hours tell you whether saved rounds convert into capacity.
A tool piloted on single-family alterations will outperform one evaluated across commercial new construction, and an unsegmented baseline lets that difference pass unnoticed when a vendor reports the after. Departments already running digital intake on structured data can compete for the demonstration funding and report honestly on what it bought them. That modernization pays off in shorter queues and auditable records whether or not a model ever sits on top of it.
Sources
- U.S. Department of Housing and Urban Development, Automated Permitting Systems Demonstration NOFO, simpler.grants.gov
- National Association of Home Builders, Home Builders Tell Congress How Permitting Roadblocks Raise Housing Costs, February 2025
- Government Technology (via Stateline), Cities Are Using AI to Speed Up Housing Project Permitting
- Government Technology, Honolulu Launches AI-Assisted Fast-Track Permit Review
- City of Fairfax, Virginia, Permit Review Metrics
- Spruce, The Last Twenty Percent Problem in GenAI Solutions, November 2025
- Spruce, Permitting Automation for Municipal Construction
