Permit review should start with completeness, not with a human opener

Backlogs start before anyone reads the file

In Boston, median discretionary permitting runs 7.5 months, according to the White House Council of Economic Advisers. It runs 13 months in Los Angeles, 16 months in Seattle, 30 months in New York City, and 33 months in San Francisco. Those waits mean months of carrying costs, financing pressure, and delayed starts before a shovel reaches the ground.

The problem is not that reviewers work too slowly. It is that the work is badly staged. Applications often arrive incomplete, inconsistent, or split across PDFs, drawings, GIS layers, parcel records, zoning rules, and prior approvals, so staff spend their time asking for missing pieces before they ever reach the merits of the permit. According to the White House Council of Economic Advisers, review is often serial, manual, and iterative, with repeated correction cycles that add weeks or months before a file reaches a human reviewer.

The fastest way to cut permit delay is not to ask humans to read faster. It is to stop incomplete packages from entering the human queue in the first place. A software check of completeness, with a person still making the final call, changes the work from chasing missing forms to deciding whether the file is ready.

That matters because the costs compound. The White House Council of Economic Advisers says each additional month in the permitting process in Washington state adds about $4,400 to the cost of building a home, roughly 1 percent per month. NAHB says a statewide average permit delay of 6.5 months was associated with about $31,375 in holding costs per home. In New York City, the White House Council of Economic Advisers estimated that a two-year delay for mid-rise development raised per-unit cost by about $50,000. When a queue is expensive, it is not harmless to let bad submissions wait in line.

The queue is slow because the file is incomplete

Permit systems persist in this shape for familiar reasons. The process is fragmented across planning, building, fire, utilities, transportation, and records. The data is scattered across the permit case-management system, document stores, GIS, parcel records, zoning rules, prior permits, and fee systems. Agencies are usually judged on compliance and risk avoidance, not throughput, so the safest choice is often to send an application back for more information rather than accept a borderline package and move it ahead.

Applicants respond rationally to that structure. If the queue is long, they submit early or partial packages to keep their place in line and fill in the gaps later. Staff then become the completeness checker by default. That is how a document review problem turns into a queueing problem.

The scale of the drag is not small. A recent analysis of 1.1 million permits across 60 U.S. cities found that a standardized 20-unit apartment building takes 4.6 years on average to complete in San Francisco, New York, and Los Angeles, with permitting delays accounting for 30 percent to 43 percent of that gap, according to HousingWire. Across the same 60-city sample, permitting time for a standardized multifamily building rose from 0.6 years in 2000 to 1.5 years in 2023, while single-family permitting rose from 1.6 months to 4.9 months, according to Vox. This is not a niche inefficiency. It is part of the cost structure of housing delivery.

For infrastructure and commercial work, the same pattern appears in different clothes. A file that lacks an attachment, mismatches the parcel, or omits a required signoff does not need a senior reviewer. It needs a gate that can identify the deficiency before the file enters the review queue. That is the job a completeness assistant can do.

The useful system is narrow and boring

The right design is not an all-purpose planner. It is a front end that checks whether the submission is complete enough for a human to spend time on it. Government Technology and Enterprise AI Case Studies both describe municipal use cases where AI sits in front of the review process, checking required fields, documents, and attachments before staff touch the file.

A reference design is straightforward. The application arrives from the permit case-management system and document store. An extraction layer reads forms, plans, scans, and attachments. A rules engine compares the submission against jurisdiction-specific checklists, zoning rules, and required artifacts. A retrieval layer brings in parcel data, address records, prior permits, and external overlays such as floodplains or right-of-way constraints. The assistant then returns one of two outcomes: ready for human review, or incomplete with a deficiency list tied to the exact rule or missing artifact.

That is where the value sits. Not in replacing review, but in changing what gets reviewed. If a site does not match the parcel record, if a required drawing is missing, or if the wrong zoning checklist is attached, the file should bounce before it enters a human work queue. If it passes, the reviewer sees a cleaner package and spends time on judgment instead of clerical correction.

The operational logic is similar to intake triage in other regulated workflows. Human work becomes more expensive when it is spent on obvious omissions. Machines are better suited to compare a package against a checklist, check that the address matches the parcel, and flag a mismatch between the submitted form and the underlying record. The reviewer still decides, but the reviewer starts with a file that has been screened.

The architecture below shows the sequence from source data to decision.

Human control stays in the loop, where it belongs

The strongest objection is not technical. It is governance. A completeness assistant can produce false positives that send applicants away unnecessarily, or false negatives that let defective applications into the queue. It can also reflect the same inequities already baked into permitting, because incomplete submissions may come from smaller applicants, language barriers, or uneven access to consultants. If the correction path is opaque, the system simply automates frustration.

That is why traceability matters. Every flag should link back to the rule, checklist item, or missing artifact that triggered it. Human reviewers should keep final decision rights. The ruleset should be versioned so the agency can show what was checked at submission time if a permit is later challenged. And because permit packets can contain personally identifiable information, property data, and building plans, access controls and audit logging are not optional.

There is also a practical limit to what software can fix. If the jurisdiction itself has inconsistent rules, if departments do not share a common records model, or if the agency cannot connect permit, GIS, parcel, and prior-approval data, completeness screening will be less accurate. The system does not solve policy friction. It exposes it sooner and costs less to correct.

But that is still a meaningful change. Current permit operations make staff absorb the cost of incomplete packages. A front-end completeness gate pushes that burden back to the moment of submission, where it belongs. The human reviewer then opens only files that are ready to be judged.

Leaders should treat permit intake as a workflow problem, not a forms problem

For CIOs, CTOs, and operations leaders, the decision is less about housing than about process design. The same pattern shows up anywhere a high-value queue is clogged by missing documents, mismatched records, and repeated handoffs. The fix is to move validation upstream, connect the systems of record, and make the first pass machine-readable before a person spends time on it.

That usually means starting with a narrow workflow, not a broad platform rewrite. Pick one permit type, one jurisdictional checklist, one intake path, and one clear answer: ready or incomplete. Then connect the permit case-management system, document management system, GIS and parcel data, zoning knowledge base, fee status, and routing queue. If the assistant cannot explain why a file was flagged, it is not ready for production.

QueryNow takes that same approach to the intake problem. If you want the workflow gone, use Build Your AI and tell us the workflow you want gone, we build it in your environment in two weeks, and you pay $10,000 only after it meets the acceptance criteria you signed off on. If permit backlogs are being treated as normal, they will stay that way.

For teams comparing approaches, AI Governance is the right next step.