HUD Is Paying Cities to Run Plan Review on AI
In June 2026 the federal government offered building departments up to $1.5 million to put plan review on AI software. Buried in the same notice was a sentence that belongs in every applicant's reading file. HUD wrote that while these systems "are increasingly marketed to state and local governments, there is limited publicly available documentation of implementation experiences, staffing implications, and governance considerations."
That is a refreshingly candid thing for a funder to say, and it's a good sign. AI plan review is moving fast and showing real promise, and HUD is putting money behind the studies that will turn early wins into a repeatable playbook. The technology is simply ahead of the documentation right now. This grant is how the evidence base catches up.
That gap is not a reason to ignore the grant. It's the grant's most useful artifact. The three things HUD says it can't document (implementation, staffing, and governance) are exactly the three things a department has to answer before it spends a dollar on automated review. The binding constraint on whether this works is not the model's accuracy. It's whether the code is clear enough to check, the desk is staffed enough to keep humans in the loop, and the decisions leave an audit trail. The deployments that pay off will be the ones that fixed those first.
What HUD is actually offering
The program is the Automated Permitting Systems Demonstration, opportunity number PDR-2600-DC-029O, run out of HUD's Policy Development and Research office. Three million dollars total, split into six awards of $300,000 to $1.5 million each. Eligible applicants are state, county, city, township, and federally recognized tribal governments. No cost-share. Applications close July 13, 2026. The grants.gov timeline puts awards around September 1 and projects starting around October 1, though HUD has not confirmed those dates against the posted NOFO PDF.
The telling detail is what the money covers. It funds roughly three years of software licensing plus staff salaries and expenses to adopt and operate the system. HUD is not just buying licenses. It's paying for people: the reviewers, the change management, the human labor that turns a tool into a working process. An agency that thought AI replaced reviewers would not write the grant this way.
There's a reason the funding reads like a study and not a procurement. The grants.gov listing classifies the awards as cooperative agreements, the federal instrument used when the funder stays involved in the work rather than cutting a check and walking away. The listing says the demonstration will generate empirical evidence on operational performance, costs, governance needs, and potential cost savings. That is the data that, by HUD's own admission, doesn't exist yet. A department chasing the grant should understand it's signing up to be a data point, not a finished case study. Trade coverage from Smart Cities Dive, Construction Dive, and Facilities Dive confirmed the verbatim HUD quote traces to the notice itself, not a spokesperson, and reported that HUD will work closely with recipients to evaluate the results.
HUD is funding six live experiments, holding the right to study them, and writing into the award the question it most wants answered: do these systems actually work in a real building department, at what staffing cost, under what controls? The agency is not betting the answer is yes. It's buying the answer. That is an unusual thing for a funder to admit, and it's the clearest signal in the notice about how seriously an applicant should take the unknowns.
What the only public pilot data actually shows
The best evidence in the country was published two weeks before the deadline. On June 17, 2026, Seattle released the results of a controlled pilot in which an AI tool screened real applications and the city's reviewers screened the same ones, then compared the two. The tool hit 87 percent accuracy on completeness checks and 92 percent on compliance checks.
Those are strong numbers. Seattle still recommended not deploying the tool for compliance. The recommendation was to move forward on completeness pre-screening only.
The reason is the part that matters. Accuracy is not the same as coverage. A tool can be right 92 percent of the time on the checks it makes and still miss corrections a human reviewer would catch. Compliance review means comparing a submission against thousands of requirements, some of them, in the city's words, "conceptually complex and likely difficult to automate."
Completeness asks whether the right documents are present and filled in. Compliance asks whether the design is legal. The first is a checklist. The second is judgment. AI cleared the checklist far better than the judgment. That split is the same one Austin departments draw between the completeness check and substantive review, and Seattle's data put numbers on it.
The most useful line in the report is a confession about the city's own code. Seattle concluded that "simplifying the City's own codes and processes would make it easier to automate both completeness and compliance checks in the future." The tool's ceiling was set by the legibility of the rules it was asked to apply. Where the code was clean, the software worked. Where the code was tangled, no accuracy figure could save it.
The part the brochures skip: staffing and governance
Honolulu is the case that shows what happens after launch. The city's permitting department fully launched an AI pre-screening tool on December 8, 2025, and the speed gains are real. The median overall permit wait fell about 40 percent, from a four-month average the prior year to about 2.5 months across the recent August-through-April window. DPP Director Dawn Takeuchi Apuna had named the target up front: "One of the biggest challenges that create permitting delays," she said at launch, "is submission of building permit applications that are incomplete and of poor quality, requiring corrections that result in additional review cycles that extend the overall building permit process." That's a completeness problem, and the tool moved it.
The gains aren't even, and the unevenness is the lesson. Residential permits got faster. They're more standardized, more checklist-shaped, the kind of submission where a completeness pre-screen has clean rules to apply.
Commercial permits did not improve, and in places got slower. They carry more conditions, more discretionary review, more of the judgment that an accuracy figure can't capture. The tool sped up the work where the code was clear and stalled where it wasn't. That is the same result Seattle reported, this time at production scale and split down the line between two permit types in one city. A department reading the grant should notice that the technology did not lift every queue equally. It lifted the legible one.
The reported headline figures are louder than the defensible ones. A pre-screen drop from five months to three days and a 71 percent cut across a 24-project trial have circulated, but they're harder to pin down, and Civil Beat flagged the trial number as one it could not independently verify. The 40 percent median figure is the one that holds, and it's the one a serious applicant should plan around.
The faster number is not the whole story. Honolulu's permitting department had 76 of 365 positions unfilled as of June 2026, a vacancy rate near 20 percent. A tool that depends on humans staying in the loop is running on a desk that's a fifth empty. HUD's grant covers staffing for a reason.
The governance gap is sharper. Honolulu lived through a permit bribery scandal in 2021. The new permitting platform lets staff selectively skip review stages, and no ethics controls were added beyond the citywide biennial training that was already mandatory. Faster software with a skippable review path and no new oversight is a control weakness with a deadline attached. That is precisely the "governance considerations" HUD said no one has documented. Honolulu is documenting it.
Where this came from, and the liability nobody priced
The grant didn't appear out of nowhere. It descends from Executive Order 14394, "Removing Regulatory Barriers to Affordable Home Construction," signed March 13, 2026. The order directed HUD to develop permitting best practices, including "capping permitting timelines and fees" and "allowing by-right development for single-family homes." It also reached into the Clean Water Act, directing review of Section 404 wetlands permitting and the stormwater Construction General Permit. The NOFO is one of the order's instruments.
One provision matters more than the timeline caps for the professionals who'll be holding the bag. The order calls for "third-party inspections and appropriate builder choice on certified entities for inspections and studies," which is private plan review and inspection in plainer terms.
When a city's reviewer signs off, the city carries the risk. When a private entity signs off, the risk moves to the private entity and the design professional whose stamp is on the drawings. Pair that with AI pre-screening and the question gets pointed. If a tool clears a submission a human would have flagged, who answers for the miss? The engineer who relied on it doesn't get to point at the software. That's the same exposure that lives in the gap between what a stamp warrants and what a standard contract covers, and automation widens it rather than closing it.
The people who build these tools say so directly. At the National Institute of Building Sciences conference in May 2026, Colin Whitlatch, CTO of construction-software firm Kahua, described the design principle: "We operate off of something called human in the loop. Because one thing is AI, with limited exception, never says, 'no.'" A system that rarely denies is a system that needs a human empowered, and accountable, for the denials. That human has to be staffed and trained, and left on the hook for the call. It all loops back to the three things HUD admits it can't yet document.
Turning HUD's three unknowns into a checklist
HUD handed every applicant a free diligence framework by naming what it doesn't know. A department should answer all three before it deploys, not after.
The first question is whether the code is clear enough to automate. Seattle's own answer was no, not yet, and it recommended simplifying the code before expecting compliance automation to work. The move is to scope the tool to completeness, the measurable checks, and keep judgment with reviewers until the rules themselves are legible. A department that hasn't cleaned up its code is buying a tool that will hit its ceiling on the first complex project, the same way Honolulu's commercial queue stalled while its residential queue cleared.
Who stays in the loop, and whether they're actually there, is the next question. AI that never says no needs humans who can, and Honolulu's 20 percent vacancy rate is the cautionary number. The grant funds salaries on purpose. A serious applicant uses that money to staff the human side of the loop rather than to thin it out, because a pre-screen that flags problems is only as good as the reviewer left to act on the flag.
The third question is whether you can reconstruct why a permit was approved. If the software can skip review stages, the audit trail has to record who skipped what and why. Honolulu's post-scandal history with no new controls is the example of what not to carry into a faster system. Before automation speeds anything up, the decision record has to be auditable, because speed without a record is just a faster way to lose track of who decided what.
Here is the part worth getting excited about. When the code is clean and the reviewers are in place, AI plan review does something permitting has never managed before. It gives an applicant a near-instant read on whether a submission will clear, and it gives reviewers their time back for the judgment calls that actually need a human. Honolulu cut its waits by 40 percent. Seattle's tool was right on more than nine of every ten checks. This is not a far-off promise. It is happening now, in real building departments.
The jurisdictions that pair that capability with clear code, a staffed desk, and an auditable record won't just permit faster. They'll set the standard everyone else copies. HUD's grant is the on-ramp, and the cities that take the prep work seriously are about to make permitting feel like a fundamentally better process. For where AI plan review already stands in practice, the cities that have already handed plan review to AI show what that looks like.