Set aside the pilots and the slide decks. There is a short list of tasks in compliance software where AI already earns its keep, and every one of them has the same shape: the model drafts, classifies, or flags — and a person decides.
Five Tasks Worth Automating
Summarizing public comments. A conditional use permit draws 600 written comments. By the next morning they are grouped by theme — traffic on the access road, noise at the property line, two comments about water rights that nobody expected. Staff still read the originals before writing the staff report, but they start knowing the shape of the record instead of discovering it page by page.
Drafting letters and reports. A code enforcement officer opens a violation case and the system drafts the notice from the case record: service address, observed condition, code section, deadline to comply. The officer fixes one sentence, deletes another, and signs. The letter goes out under a person's name, with the case file behind it — the draft just saved forty minutes of retyping things the system already knew.
Classifying documents. A certified tester emails a PDF. It gets filed as a test report against the right assembly at the right service address, with the tester's certification number read off the form. The utility clerk confirms the match instead of typing it — and when the clerk corrects one, the correction teaches the classifier.
Flagging missing fields. An application arrives without the site plan, or with a tester certification that expired in March. It gets flagged at intake — not an hour into a reviewer's afternoon, and not three weeks later in a completeness letter that restarts the clock.
Explaining status. An applicant calls asking where their permit is. The answer — resubmittal received Tuesday, plan review in progress, waiting on the fire marshal — comes in plain language, drawn from actual workflow state rather than from whoever happens to remember the file.
Each of these removes hours of mechanical work. None of them removes a decision.
Where It Doesn't Belong
The line is bright: AI does not decide compliance. It does not determine that a property is in violation, that a permit is denied, that an assembly failed its test, or that an account escalates to a shutoff notice. Those are decisions a public agency has to defend — to an appeals board, a court, or a records request — with a named human and a stated reason. The decision produced by a model is the decision nobody can explain. Agencies should refuse that position on principle, before any vendor demo, because no demo ever shows the appeal.
The Test: If This Output Is Wrong, What Happens?
Apply this question to any AI feature and the evaluation is mostly done. If a wrong draft gets fixed in review, the feature saves time at low risk. If a wrong output becomes a shutoff notice sent certified mail to an address the customer left two years ago, the feature is making decisions — whatever the brochure calls it. Drafting versus deciding is the whole distinction, and the test finds it every time.
The Records Have to Be Right First
Every use case above assumes the boring foundation: accurate records, statuses that mean one thing, and a review step that leaves evidence. If a status of Passed sometimes means the test passed and sometimes means a clerk closed the record to clear a queue, the AI's plain-language answer will be confidently wrong — and confident wrong output, mailed on letterhead, is the most expensive kind.
So skip the question of whether to use AI. List the drafting tasks in your workflow — the letters, the filing, the intake checks — and ask whether your review step would catch a bad draft before it leaves the building. If yes, automate the draft. If no, fix the review first.