When I ran shed and garage permits as a permit technician, the question that ate the week was never a hard one. It was some version of "does this one need a permit?" Our department absorbed roughly two hundred of those a week, all from inside our own company, divided among whoever had capacity, each taking three to fifteen minutes.
None of that time was analysis. It was locating. Which municipality. Which department inside it. Which page on which website. Which PDF, and which revision of the PDF.
A full requirement check, the kind where you write down setbacks and the submittal list for a real address, ran about twenty minutes on a good day. Good days were not the norm. Some of them turned into multiple days of waiting on a callback from an office that answered questions certain hours on certain days, where the one person who could answer the question was frequently not in.
That specific work is what AI is good at now. It is worth being precise about which part, though. The part it cannot do is the part that decides whether your project gets built where you drew it.
Retrieval at a scale a person cannot matchRetrieval at a scale a person cannot match
Machines genuinely do four things well in permitting, and every one of them is a version of reading fast.
Resolving the authority. Given an address, working out which authority having jurisdiction actually issues the permit is a real problem in the Midwest. A parcel with a village mailing address can sit in unincorporated county. Building, zoning, and health can be three separate desks with three separate answers. That is a data problem, and data problems are where software wins outright.
Reading everything at once. A jurisdiction's requirements are scattered across a fee schedule, a submittal checklist, a zoning table, portal help text, and a counter handout. A person opens those one at a time and gets tired. A model opens all of them together.
Pulling structure out of prose. Turning "accessory structures shall be located no less than five (5) feet from any interior side lot line" into a setback value attached to a structure type is the tedious middle of the job, and it is exactly the shape of task language models handle.
Checking a packet against a list. Comparing what you have against what the department expects is mechanical. Missing signature, missing survey, missing detail sheet.
At Permitech, an AI research pass on a single address takes about three to five minutes. A full requirements set for that address and scope lands in roughly half an hour, across the 7,000-plus jurisdictions in Illinois, Wisconsin, and Indiana. Set that against twenty minutes per check on a good day, plus the callback wait, and the arithmetic is not subtle.
The five places retrieval runs outThe five places retrieval runs out
Now the honest half. There are five failure modes here, and reading faster does not touch any of them.
An ordinance that is genuinely ambiguous. Some code language does not resolve on the page. Whether a roofed patio reads as an accessory structure or an addition, whether a landing counts toward the deck area threshold. The document supports two readings. A model that produces one confident answer has produced a guess wearing formatting.
A checklist that is three revisions behind. Departments change internal practice faster than they change their website. The PDF sitting on the permits page can be older than the standard the reviewer applies. A model reads that PDF accurately and reports it accurately, and is still wrong about today. This is the failure that hurts most, because it fails silently.
Unwritten local practice. Every counter carries knowledge that exists in no document: which reviewer wants the site plan dimensioned a particular way, that zoning sign-off happens before intake rather than after, that a certain file always routes to the county. You learn it by calling, or by getting it wrong once.
A judgment call. Whether a hardship meets the standard for a variance is a discretionary decision made by a board after a hearing. Software can tell you the standard. It cannot tell you how your specific lot shape reads to five people in a room, and any tool that claims to is selling a prediction as a fact.
Accountability. When a requirement turns out to be wrong, somebody has to own it, call the department, and fix the file. No model does that. That is not a capability gap that shrinks with a better model, it is a structural one.
Most permit failures are not lookup failuresMost permit failures are not lookup failures
Here is the part that reframes the whole question. Better research fixes research problems, and research problems are a minority of what actually goes wrong.
Across more than 220 building permits I submitted as a permit technician and can still account for in records, 37 were logged as cancelled. The breakdown is not what you would guess from reading permit software marketing.
One clarification before the chart, because the distinction matters. Cancelled is not a synonym for denied. Only part of that number was a building department saying no. Several were customers who could not produce a required document, or who could and simply would not go get one. A few were projects that turned out not to need a permit at all, which is a good outcome recorded as a dead file.
Why 37 submitted permits were logged as cancelled
More than 220 submitted permits, residential sheds and detached garages. Cancelled covers every file that closed without an issued permit, which includes customer-side withdrawals, not only department denials.
Source: Permitech analysis of the founder's permit technician case records
Analysis by Permitech
Free to cite with attribution to Permitech and a link to this page. Republishing the chart or the underlying table without attribution is not permitted.
View the data
| Category | Value (cases) |
|---|---|
| Customer self-filed or declined service | 11 |
| Placement problem | 10 |
| Extra permit requirements surfaced | 5 |
| No permit actually required | 4 |
| Customer would not provide a plat of survey | 4 |
| Septic complication | 1 |
| Wrong address | 1 |
| Cancelled after the permit issued | 1 |
Sort those by what a faster lookup would have prevented and the chart mostly empties out. Eleven cases were customers deciding they did not want the service, which is a sales outcome. Four were projects that needed no permit at all, and those are the ones a good research pass genuinely catches before anyone spends money.
The rest are workflow. Placement, at ten cases, is the largest true permit failure in the log, and knowing the setback number is not the same as knowing where the lot line runs. That distance gets measured off a fence, a curb, or a hedge, and the reference line is the part people get wrong. Four more cases died because the customer would not produce a plat of survey, which is a document problem no amount of retrieval solves. Five had additional requirements surface mid-process, which is scope discovery, and it is closely related to why files come back from review.
What a verification pass is actually forWhat a verification pass is actually for
A human permit technician verification pass at Permitech runs one to four hours after the AI research completes. That number surprises people who expect software to have eliminated it. It exists because of the list above.
- Confirm the resolved jurisdiction against the parcel, not the mailing address
- Call the department when a published document and an observed practice disagree
- Check whether the submittal checklist on the website is the one currently in use
- Flag ambiguous ordinance language instead of resolving it silently
- Identify septic, well, floodplain, or county health reviews that sit outside the building department
- Pin down which documents the applicant still has to produce
- Put a name on the answer, so there is somebody to call when it is wrong
That last line is the whole argument. Everything above it is work; that one is responsibility.
Why we do not sell you an AI permit technicianWhy we do not sell you an AI permit technician
There is a version of this product that would be easier to market. Hire an AI permit technician, it files your permits, done. We are not selling it, because it does not exist yet.
Autonomous AI permit technicians are on the Permitech roadmap. What ships today is AI research paired with human permit technicians who verify and fulfill. A permit is a legal instrument. Somebody signs the application, somebody is responsible for the inspection sequence, and somebody answers when a reviewer calls. Those are accountability roles before they are labor roles.
Anyone claiming full automation today is either redefining the word or has not sat at the counter. Compare it against how expediters and managed permit services are actually structured and the distinction gets clearer fast.
The twenty minutes, not the technicianThe twenty minutes, not the technician
The win was never replacing the permit tech. Trade a good one for a model and you lose the phone call, the counter relationship, and the person who knows the reviewer's habits. You keep the reading, which was the cheap part all along.
A better trade exists. Take the twenty minutes of lookup, the callback wait, and the seventy-five internal questions a week, and hand all of it to something that does it in four minutes. What is left over is judgment. Where the structure actually sits. Whether the survey shows current conditions. Which department gets the file first.
That is the shape of a Permit Package: AI research on your address and scope, a verified requirement set on top of it, organized in a Permit Workspace with a Permit Application Reference Sheet holding the answers the application asks for. You copy those onto the municipality's form. If you would rather not touch it at all, Permit Concierge runs the filing.
The permit technician does not go away. They stop being a search engine.
