Julia Richman waited nine months. Not for a building permit to be approved. For someone at the city to look at it. She had submitted plans to renovate her 115-year-old home in 2022 and spent three-quarters of a year in a queue so deep that no human reviewer had opened the file. When Denver's planning department finally adopted AI-assisted plan review this spring, the pitch was that technology would prevent stories like Richman's from happening again. What nobody mentioned was that the documents the AI would be reading were, by any measurable standard, written to be misunderstood.
A DOE-funded study published June 16 in Scientific Reports scored roughly 300 state, county, and city permitting documents on a 1-to-5 clarity scale using an Energy Language Model fine-tuned specifically for regulatory text. Local permitting documents averaged 1.8 out of 5.
Read that number again.
That score means more than half of what a builder, architect, or homeowner encounters in a municipal permitting document is ambiguous, internally contradictory, or references other sections that themselves reference other sections in a chain of bureaucratic indirection that would make a tax attorney blink, and that no amount of careful reading by a patient applicant with a highlighter and a legal dictionary can fully resolve. And the response from governments, increasingly, is not to rewrite the documents. It is to purchase multimillion-dollar AI platforms to parse them faster.
Denver Bets $4.6 Million on Reading Comprehension
In March 2026, Denver approved a five-year, $4.6 million contract with Clariti, the company behind CivCheck, a platform branded as "Guided AI Plan Review." CivCheck screens permit applications before human reviewers touch them, flagging missing documents, incomplete fields, and code references that don't resolve. It is, in essence, an expensive proofreader for an application process that currently passes first-round review only 30 to 37 percent of the time.
Denver's planning department was cut by 59 positions for 2026, leaving 251 staff to handle a city that simultaneously enacted a 180-day shot clock for permit decisions, with a $10,000 refund if the deadline is missed. The math is straightforward: fewer people, hard deadline, financial penalty. Automate or fail. Since 2023, single-family and duplex permit processing time has dropped 45 percent. CivCheck is supposed to push that further.
But consider what CivCheck is actually doing. It is not reviewing architectural plans for structural integrity or code compliance in any engineering sense. It is checking whether the applicant filled in the right boxes, attached the right documents, and referenced the correct municipal code sections, sections that, per the DOE study, score below 2 out of 5 for clarity. CivCheck is compensating for a user interface problem by adding a sophisticated interpreter between the applicant and a code that the code's own authors could not write clearly.
Five Cities, Same Pattern
Denver is not alone in this approach, and the pattern is spreading with a speed that suggests nobody is asking the upstream question.
| City | AI Tool | Status | What It Does |
|---|---|---|---|
| Denver, CO | CivCheck (Clariti) | $4.6M contract, 5 years | Pre-screens applications, flags missing docs |
| Honolulu, HI | CivCheck | Residential live Dec 2025 | Same platform, commercial by mid-2026 |
| Seattle, WA | Custom (executive order) | Full rollout expected 2026 | AI pilot on all development applications |
| Austin, TX | Archistar | Active pilot | Zoning review automation |
| Naples, FL | Blitz AI + CityView | First in Florida | Automated plan review |
Baltimore and Corona, California, are running their own implementations, and Seattle's mayor signed an executive order directing AI integration across all development applications, creating a wave of procurement that stretches from the Pacific Northwest to the mid-Atlantic without any two cities coordinating their approach or sharing lessons learned from the deployments already underway. Each addresses the same symptom: applications take too long because reviewers spend too much time deciphering what the applicant meant, what the code requires, and whether the two actually align.
None of these deployments include a project to rewrite the permitting codes themselves. Not one.
The Altamonte Springs Cautionary Tale
Altamonte Springs, Florida, piloted AutoReview.AI in 2023 for automated plan review, a smart procurement move that checked every box a city council could want. AutoReview.AI no longer exists as a company. The unclear codes remain.
This is the vendor risk that nobody in the current procurement wave seems to be pricing in. A five-year, $4.6 million contract assumes the vendor survives for five years, and construction-tech has a graveyard deep enough to give any procurement officer pause: Katerra burned $2 billion, Veev consumed $647 million, and a string of smaller startups took municipal pilot money and evaporated. When the AI vendor disappears, the city is left with the same 1.8-out-of-5 documents and no institutional capacity to interpret them, because the human expertise was cut to fund the AI contract.
Across the Atlantic, a Different Approach (Slightly)
On June 17, the UK government announced two AI planning tools built with Google DeepMind, Google Cloud, and Faculty under an £8.2 million contract. One of them is genuinely interesting.
APD, or Augmented Planning Decisions, aims to halve householder planning application times from eight weeks to four, starting with alpha trials in Barnet, Camden, and Dorset councils. Householder applications make up 70 percent of all planning submissions in England, roughly 350,000 per year. RIBA reports that 80 percent of architectural practices are experiencing significant project delays from planning backlogs, and more than 10 percent have abandoned projects entirely because the wait killed the economics.
The second tool, Extract, does something the American deployments have not attempted. It converts decades-old planning documents into structured digital data, replacing what the UK government estimates at 255 hours of manual document work per council. Extract is essentially a rewriting tool. It does not just parse ambiguous text faster. It transforms the text into a machine-readable format that, as a side effect, forces clarity on the underlying requirements. Twenty local planning authorities have trialed it, and it is now available to every council in England.
Minister Ian Murray summarized the problem with an economy of language that American policy documents rarely achieve: "Planning officers shouldn't be spending hours digging through decades of paper records."
He's right. But neither should an AI. That is the distinction the American approach keeps missing.
The Math Nobody Will Do
Here is the question that I cannot find a single municipality asking publicly: what would it cost to rewrite the permitting code itself?
Denver's permitting code runs to several hundred pages of cross-referenced ordinances, zoning overlays, and legacy provisions that have accreted since the city's founding. Rewriting it would require legal review, public comment periods, council votes, and coordination across departments that have historically treated their code sections as sovereign territory. A conservative estimate for a comprehensive permitting code rewrite, based on what smaller cities have spent on similar modernization projects, is $500,000 to $2 million over 18 to 24 months.
Denver chose to spend $4.6 million over five years on AI that reads the existing code. Consider the alternative. If clarity rose from 1.8 to even 3.5 out of 5, first-round approval rates would likely improve by default, because applicants could understand what was being asked of them. That improvement would compound every year without licensing fees, vendor lock-in, or the risk that your AI partner becomes the next Altamonte Springs footnote.
I ran a rough calculation. At Denver's current 30-37 percent first-round approval rate, roughly 63 to 70 percent of applications require at least one resubmission cycle. Each resubmission costs the applicant two to six weeks and, conservatively, $1,500 to $4,000 in architect revision fees, re-filing costs, and project carry expenses. For a city processing 15,000 residential permits annually, that is approximately $14 million to $42 million in annual friction costs borne by builders and homeowners. Reducing rejections by even 20 percent through clearer codes would save the private sector $2.8 million to $8.4 million per year, far exceeding the one-time cost of a code rewrite.
Housing Starts Just Fell 15 Percent. This Matters Now.
The Census Bureau reported on June 17 that housing starts dropped 15.4 percent in May 2026, falling to 1.177 million units annualized. Single-family starts hit an eight-month low. NAHB projects the national housing shortage at approximately 1.2 million homes. The construction industry is short an estimated 250,000 workers every month, and that labor gap adds roughly two months to average building timelines, according to the Home Builders Institute.
Every unnecessary permitting delay compounds into this shortage, and the compound interest is brutal. A builder waiting an extra month for a permit that was rejected because the code's setback requirements contradicted its own cross-referenced zoning overlay is a builder not starting foundation work. Multiply that across thousands of jurisdictions, and the permitting clarity problem is not a bureaucratic inconvenience. It is a meaningful contributor to the housing crisis.
What You Should Actually Do
If you are a builder or developer working in a jurisdiction that recently adopted AI permitting tools, three things are worth knowing.
First, AI pre-screening does not change what the code requires. It changes how quickly someone tells you that your application is incomplete. If the underlying code is ambiguous about whether your project needs a site plan or a survey, the AI will flag the omission faster, but it will not resolve the ambiguity. You will still end up in a meeting with a planner who interprets the code differently than the last planner you spoke to.
Second, look at the DOE scoring framework from the Nature paper. It provides a quantitative methodology for evaluating the clarity of any jurisdiction's permitting documents. If your city scores below 2 out of 5, you have data to bring to a council meeting. "Our permitting code is rated F by a Department of Energy research framework" is a more effective argument than "the process is slow."
Third, check whether your AI vendor has been operating for more than three years, has municipal clients that have renewed contracts, and has a data portability clause in the agreement. If the vendor disappears, you want your city's data and configurations to survive the transition, not locked in a proprietary format that dies with the company.
Limitations of This Analysis
The DOE study examined EV charger permitting documents specifically, not residential building permits broadly. The 1.8-out-of-5 clarity score applies to the studied corpus and may not transfer directly to general construction permitting, though the scoring framework is designed to be generalizable. Denver's CivCheck deployment is weeks old, and no independent outcome data exists yet to validate their 80 percent first-round approval target. The UK tools are in early alpha with three councils, and national rollout in 2027 remains aspirational. My cost estimates for code rewriting and friction costs use reasonable ranges but have not been independently audited. No head-to-head comparison of AI permitting tools exists in the public record.