Your Next AI Plan Review Officer Is Already in the Building Department

Who Signs the Approval When the Model Flags a Setback?

A residential permit sits in the queue for eleven weeks. A vendor tool finally runs it in ninety seconds and returns forty-one findings: nine real code violations, a dozen citations to a superseded edition, and twenty that are the model misreading a dimension line. Somebody has to sort that before an approval leaves the building under a municipal seal. That sorting is the entire job.

The seat owns four things. It owns configuration, meaning which adopted code edition and which local amendments the tool checks against, because a checker aimed at the wrong edition produces confident nonsense at volume. It owns adjudication, meaning a written disposition for every finding: upheld, overruled, or referred to a discipline reviewer. It owns the record, meaning the version of the ruleset and the model that ran on the day, kept so a decision can be reconstructed a year later when it is challenged. And it owns the approval itself, which stays a human act by a named person.

Why the seat is appearing now is a supply problem meeting a procurement wave. Honolulu's Department of Planning and Permitting is deploying AI across permitting, including a CivCheck pilot on residential permits, with Director Dawn Takeuchi Apuna citing roughly 25 percent vacancy among engineers and plan reviewers. Vendors including CivCheck, since acquired by Clariti, along with Archistar's AI Precheck, OpenGov, CodeComply and BlitzPermits all sell automated plan review into building departments where a human reviewer still signs ¹. Departments are not buying these tools to replace reviewers. They are buying them because a quarter of the reviewer seats are empty and the queue keeps growing.

One caution about the title. This is a function being deployed rather than a settled job noun. Search a public-sector board for "AI plan review officer" today and you will find little; search for plans examiner, plan review supervisor, or building plans engineer at a department that just bought a checking tool and you will find the work. Write the duty into an existing classification rather than waiting for the market to name it, and expect the noun to move under you for a couple of years.

The adjacent seat worth distinguishing is the one that validates the tool itself against a standard before anyone relies on it. That is closer to what a GxP AI validation specialist does in a regulated manufacturer, and if your jurisdiction expects a formal acceptance test of the checker, do not quietly fold that into the reviewer's caseload.

Which Tells Separate a Real Reviewer From One Who Defers to the Tool?

The failure mode is not a reviewer who rejects the software. It is one who approves whatever it says because the queue is long and the tool sounds certain. Every tell worth testing measures the same thing: does this candidate treat a machine finding as a claim requiring evidence, or as an answer. Run the test on paper, in the room, with a real drawing set.

Give them a marked-up sheet and a printed list of tool findings you have already salted with three wrong ones, and watch what happens.

• They open the code before they open the finding. A real examiner reaches for the adopted edition and the local amendment sheet. Somebody who argues from the tool's citation alone has already inverted the authority.
• They can say which edition is adopted here, and hesitate about it. Adoption cycles vary by jurisdiction and amendments are local. Confident recitation of a national model code with no mention of amendments is a tell in the wrong direction.
• They overrule something and write one clean sentence explaining why. Ask for the disposition in writing. If the reasoning is "the tool is wrong," that will not survive an appeal. If it names the section and the drawing detail, it will.
• They ask what happens to the finding they overruled. The strong ones want to know whether it feeds back to the vendor, whether it recurs, and who tracks the pattern. That instinct is the difference between a reviewer and a maintainer of the review system.
• They ask about the applicant. Someone who has stood at a counter thinks about the homeowner waiting eleven weeks and about what a rejection letter has to say to be actionable. That is not softness. A finding nobody can respond to becomes a resubmittal and another eleven weeks.

The performed version is fluent about the technology and vague about the code. It will name model families, describe a workflow diagram, and then be unable to tell you whether a stair riser dimension on the sheet in front of it complies. Hire the other way around: code fluency is the scarce half, and tool operation is learnable in a month.

Which Backgrounds Produce This Person, and How Did They Get Good With AI?

The obvious feeder is your own plan review bench: certified plans examiners, building inspectors with a review rotation, and licensed structural or fire protection engineers already doing discipline review. They arrive knowing the code and the counter, and the only new skill is treating software output as a set of assertions to be tested. That is a shorter distance to travel than teaching a technologist to read drawings.

The unexpected feeders are better than they look. BIM and VDC coordinators from the design side have spent years running clash detection, which is the same motion: an automated tool emits hundreds of findings, most of them noise, and a human decides which are real and writes the disposition. Permit expediters and code consultants have argued the other side of the counter and know exactly which rejection letters are useless. Accessibility specialists who have run ADA plan review carry an unusually strong habit of citing the specific provision rather than the general principle. And a records or GIS analyst inside the department often understands the data plumbing better than anyone, though that person needs a code-certified partner rather than the signature.

How the good ones got good with AI is worth asking directly, because the answer is diagnostic. The pattern to listen for is somebody who has already been running a checker or a general assistant against their own work and keeping score. They will tell you they asked a model to summarize an amendment and then found the two clauses it dropped. They will describe running the tool on a set they had already reviewed by hand, precisely to see where the two disagreed, and keeping the disagreements. They will have a rough sense of the tool's failure shapes: dimension misreads, superseded citations, confident answers about conditions that are simply not shown on the sheet.

What you do not want is either extreme. A candidate who has never used these tools will misjudge which findings are cheap to check and which are expensive, and will either rubber-stamp or drown. A candidate who trusts the output will approve a set that a homeowner's attorney later reads line by line. The one to hire uses the tool constantly and distrusts it specifically, and can describe the distrust without being prompted.

Recruit Where Code Certification and Automated Checking Already Overlap

Start inside. In a department running a checking pilot, one or two examiners are already doing this work informally and complaining about it in the right way. Promoting from there and backfilling behind them is faster than an outside search and it preserves the local amendment knowledge that takes a new hire a year to rebuild.

Outside, recruit where certification and practice already concentrate. International Code Council certification cohorts and local ICC chapters hold the plans examiner and building official credentials this seat needs. State and regional building officials associations, such as the California Building Officials group and its counterparts in other states, run the conferences where automated review is currently being argued about rather than announced. Local AIA chapters and the National Institute of Building Sciences reach the design-side people who have run model checking against a code ruleset. Vendor user communities are the narrowest and most useful venue of all: the departments already piloting these tools have staff who have made every mistake you are about to make.

Search on duty rather than title. Set alerts for plans examiner, plan review supervisor, building plans engineer, code compliance specialist and permit services manager at jurisdictions that have recently procured a checking product, since a procurement notice is a public document and a hiring signal. The people you want are frequently in seats that predate the software by twenty years.

Screen on artifacts, which in this field are unusually easy to get. Ask for a redacted correction letter the candidate wrote, or a plan review comment set. Read it before the interview. A correction letter shows you whether the person cites a section, describes the deficiency in terms an applicant can fix, and separates a real violation from a preference. That single document predicts performance in this seat better than any credential list, and it is the same skill the machine findings will demand at ten times the volume.

How Do You Close This Hire, and Does the Work Sit at the Counter?

Close on authority and caseload before pay, because in a public department pay is often the least movable part and authority is free. Say in writing that the reviewer may reject the tool's output, that the rejection is recorded rather than escalated, and that adjudicating machine findings counts against caseload rather than sitting on top of it. A pilot that quietly doubles the queue while promising relief is the reason the last person left.

On compensation, the honest answer for a category this new is a band rather than a number. There is no wage series for an AI plan review officer, so the seat hires against the senior plans examiner or plan review supervisor band in your existing classification schedule, with the AI duties argued as a reclassification or a working-out-of-class differential rather than a new schedule. That framing matters practically: a public department can usually move someone a step or two inside an existing class far faster than it can create one. The broader market pressure is real and measurable at the macro level. PwC's 2026 AI Jobs Barometer, reading roughly one billion job ads, found an average wage premium of about 62 percent for roles requiring AI skills ². Your certified examiner who can also run and overrule a checking tool is now being priced by private code consultancies and vendors, not only by the jurisdiction next door.

What kills the offer is predictable. A reporting line through the office that bought the software, so the reviewer's job becomes proving the purchase worked. A pilot with no defined exit criteria, which reads as a permanent unpaid evaluation. And a hiring process measured in months while a code consultancy closes in two weeks. Public departments lose these candidates on process length more than on salary.

On location, most of this seat travels. Reading drawings, running the checker, and writing dispositions are desk work, and remote or hybrid plan review is now ordinary practice in many jurisdictions. Three parts do not travel. Counter duty and applicant meetings still land in person in most departments. Field verification, when a review question can only be settled by looking at the site, is on-premise by definition. And some jurisdictions impose residency or in-state licensure requirements on staff who sign approvals, which is a question for your own counsel and civil service rules rather than something to infer from another city's posting. Write the split into the job description. A reviewer who expected remote and got a counter rotation resigns in the first quarter.

One staffing pattern worth borrowing from construction: the departments that make this work pair the reviewer with someone who owns the data and the schedule, much as a digital project controls analyst does on the delivery side, so the reviewer spends the day on code judgment rather than on spreadsheets about the queue.

See how it works >