The Permit Packet That Keeps Coming Back

A permit packet keeps coming back because the same small misses get past the people who put it together, and every round of corrections costs about a month. Cities are now running AI over packets before a reviewer opens them, and the misses it catches are the repeat ones. A contractor can run that same check on their own packet first.

Here is how it usually goes. A remodel, an addition, or an ADU gets drawn, the packet goes in, and four weeks later it comes back with a correction list. The smoke alarm locations are not shown. The lot coverage number does not match the site plan. The building height is missing from an elevation. Somebody fixes it, it goes back in, and the clock starts again. The crew you planned to put on that job goes somewhere else, or sits.

Nobody did anything careless. The designer knew the code. The office knew the process. The problem is that the list of things this particular city always asks for lives in a few people's memories and a stack of old correction letters, and nobody checks the packet against it before it goes out.

Why does a permit packet keep coming back?

Because each review cycle catches only what it catches, and each cycle takes about a month. Seattle's Department of Construction and Inspections reports that a typical middle housing application (townhouses, single family homes, ADUs) goes through 2.6 intake screening cycles and 3.7 plan review cycles before approval, and that each cycle takes about one month of calendar time.

That means a packet that comes back twice at intake has already cost the job two months before a plan reviewer has looked at the design. For a remodeler or general contractor, that is two months of a signed job you cannot start and cannot bill.

What did Seattle and Honolulu learn when AI read the packet first?

They learned that a machine can catch most of the routine misses before a person ever sees the packet. Both cities tested the same tool, CivCheck, which reads plan sets and runs fixed checks against the code.

In Seattle's pilot, the vendor ran 29 past residential applications through the tool and the city compared its findings with what its own intake reviewers had flagged. Of the 55 rejection reasons the tool was set up to check, it caught 48, or 87 percent. In plan review, reviewers agreed with its pass or fail calls 91.7 percent of the time across 1,262 checks. The city then modeled what would have happened if applicants had fixed those errors before submitting.

Days to get through intake, Seattle

Median calendar days to pass intake screening, 29 past residential applications.

• As it actually wentNo pre-check, all manual review

  54.5 days2.55 cycles

• With the pilot checks34 completeness checks, errors fixed first

  29 days1.75 cycles

• With every intake checkProjected, all completeness errors fixed first

  8 days1 cycle

Honolulu reported bigger numbers from live use. In the first quarter of 2026, residential permits that went through AI pre-screening averaged 1.4 plan review cycles instead of 3.4, 7.7 corrections instead of 23.5, and 32.5 days to a decision instead of 73. On September 1, 2026, the city made pre-screening required for new homes, additions, alterations, and ADUs, and it now rejects those applications if they skip it.

Read both of those carefully. Seattle's time savings are modeled, not observed, and they assume every applicant fixed every flag. Honolulu's comparison is 19 pre-screened permits against 17 that were not, from before the tool was required, so the two groups were not assigned at random. That is an association, not proof that the tool caused the difference. Both point the same way, and both are small.

Which corrections show up again and again?

The same handful, over and over. In the Seattle sample, the most common zoning correction was maximum lot coverage, and the most common building and fire corrections were carbon monoxide alarms and smoke alarms, each missed on 8 of the applications the city checked.

Zoning

Applications with the error, and the error rate

• Maximum lot coverage

  9 · 75%

• Trees: caliper inches

  8 · 67%

• Building height

  6 · 50%

• Front yard

  6 · 50%

• Floor area ratio

  5 · 42%

Building and fire

Applications with the error, and the error rate

• Carbon monoxide alarms

  8 · 44%

• Smoke alarms

  8 · 44%

• Fire separation distance

  5 · 28%

• Required heating

  3 · 17%

• Stairway handrail

  3 · 17%

None of those are hard. Showing alarm locations on a floor plan, stating lot coverage, and dimensioning the rear yard are things every designer knows to do. They get missed because nobody reads the whole packet against the whole list one last time before it goes in. That is a reading job, and reading against a list is what AI does well.

This is the same idea I wrote about in your business runs on what people remember. The corrections your city always asks for are a process. Right now it lives in memory. Written down once, it runs the same way on every packet.

What can a contractor check before submitting, and what still needs a person?

A machine can check whether something is there and whether a number meets a rule. A person decides what the site and the project actually need.

The machine checks

• Is every sheet there?

  Site plan, floor plans, elevations, sections, and the forms your city requires.

• Is every callout shown?

  Smoke and carbon monoxide alarm locations, heights on elevations, tree sizes.

• Does the number meet the rule?

  Lot coverage, setbacks, floor area ratio, building height.

• Do the documents agree?

  The application, the site plan, and the floor plans state the same numbers.

A person decides

• Should this requirement be waived?

  A survey for work confined to an existing basement, for example.

• What does the site actually need?

  Slopes, soils, trees, and neighbors the drawings do not fully show.

• How do we handle an unusual design?

  Edge cases a fixed rule was not written for.

• When do we push back?

  A flag that is right about the rule and wrong about this job.

That split matters more than the tool. Seattle found that 35 of the 90 rejection reasons in its sample were outside what the tool was set up to check, and the city wrote that its own codes do not always translate into clear machine logic. The Florida League of Cities, in its August 2026 buyer's guide for cities, puts it plainly: in every tool, AI flags issues and points to the rule, but a person still makes the decision.

What happens when the AI is stricter than the reviewer?

It flags things a person would have waived, and someone has to know when to push back. In Seattle, the tool flagged four applications as needing a topographic survey that the city's intake staff had exempted. In one case the site had steep slopes, but the work was confined to an existing basement, so the reviewer checked with the geotechnical team and waived the survey.

The machine was not wrong about the rule. It just did not know the exception, because the exception was never written down. I made the same point about AI agents in the loop is the easy part: the machine can try and check, but knowing what good looks like on this job is the part someone has to supply. On a permit, that person is your designer or your project manager, and their judgment calls should be recorded so the next packet benefits.

Why does this matter now?

Because cities are moving fast and the rules for applicants are changing. Honolulu went from a pilot to a requirement in less than a year. Florida's HB 803, which took effect July 1, 2026, tightens permit review deadlines and creates a five business day review window for certain small residential permits, according to the Florida League of Cities, and the League's guide notes that tools catching incomplete applications on day one help departments meet those deadlines. In Hernando County, the county reports that an AI review tool helped clear a backlog of about 6,000 single family applications after the 2024 hurricanes. That figure comes from a vendor case study, so treat it as the county and vendor's account.

If your jurisdiction is not there yet, you do not have to wait for it. The check works the same way on your side of the counter.

The honest limits

This check only knows what is on your list and in your packet. If your city adds a new requirement and you have not seen it in a correction letter yet, the check will not catch it. AI can also misread a dense plan sheet, which is why it flags rather than fixes, and why a person reads every flag.

It does not replace your designer, your engineer, or your relationship with the permit counter. It will not argue a code interpretation. And the numbers in this piece are early. Seattle's results come from a controlled test where the vendor configured the tool, on a small sample of residential projects, with time savings modeled rather than measured. Honolulu's comparison is small and not randomized. Most of this data covers residential building permits. Honolulu exempts general electrical, plumbing, and solar work for now, so if your shop mostly pulls trade permits, your own correction letters are the better guide.

How would ClearOak address it?

We would start with your correction letters, not with software. I would sit down with whoever assembles your permit packets and pull the last 10 to 20 correction letters from the jurisdictions you work in most. Those letters are the real list of what your cities always ask for.

From there, the work is five steps:

• Build the corrections list. Group every correction from those letters by jurisdiction and by type: missing sheet, missing callout, number that does not meet a rule, judgment call. That becomes a one page checklist per city your team agrees on.
• Decide who owns which call. The checklist marks which items a machine can confirm and which ones need your designer or project manager, like waivers, site conditions, and unusual designs.
• Read the packet against the list before it goes in. An AI check reads the plan set and the application, compares them to that city's list, and returns a short set of flags. It does not change a drawing or submit anything.
• Run it next to your people on a few packets. Your team keeps doing what it does, and we compare the flags to what came back from the city. That shows where the list or the check has gaps before you rely on it.
• Hand it over, and keep the list alive. Every new correction letter adds a line to the list. Your team owns that, and we train them to update it without calling us.

ClearOak leads this as operators first. We bring in the right builders for the connection to the tools you already use, so the check fits how your office already puts a packet together.

What you get back is time and revenue. Your office stops chasing the same corrections on every job, and signed work starts and bills sooner.

How do you start with your next permit?

Pull the correction letters from your last five permits in one jurisdiction and write every item on a single page. Before your next packet goes in, have someone who did not assemble it read it against that page. You will see how many of your corrections were repeats, and you will have the first version of the list that everything else is built on.

If you want help deciding which part of this is worth automating, what not to automate is a good place to draw that line.

Where ClearOak comes in

ClearOak & Co. is a software and operations partner for trades businesses. We find the process costing you the most, build what runs it on your real data with the right builders for the job, and train your team to own it.

If your permits keep coming back with the same corrections, schedule a call and we will look at your last few correction letters together.

Prefer email? Reach me at john@clearoakconsulting.com and tell me which correction your city sends back most often.

Sources

• CivCheck AI Pre-Screening Pilot Report, Seattle Innovation and Performance with the Seattle Department of Construction and Inspections, June 2026 (proof of value on 29 intake and 57 plan review applications; vendor configured the tool)
• City earns national AI 50 award for innovative permit review technology, City and County of Honolulu Department of Planning and Permitting, 2026 (Q1 2026 figures from the vendor's performance report; 19 pre-screened permits compared with 17 that were not)
• New AI tool now required for some building permit applications on Oahu, Hawaii News Now, September 1, 2026
• Buyer's Guide: AI Tools for Building Permitting, Florida League of Cities, August 2026 (Hernando County figure from a vendor case study)
• CS/CS/HB 803: Building Permits and Inspections, Florida Senate, Chapter 2026-63, effective July 1, 2026