Guide

Where human approval belongs in a municipal AI workflow

In a municipal service the question is not whether people stay involved but where their attention is needed. These principles come from Liba’s work with municipal and utility services.

Separate reading, rules and decisions

A sound design gives each part one job. AI reads documents and extracts fields. Rules approved by the municipality determine what the checks mean. An employee approves. When the three are mixed, it becomes hard to explain why a case was decided the way it was.

Approval points to keep

• Every approval and rejection in the early stages. The system recommends and an employee confirms. Automatic execution of routine approvals, if it happens at all, comes after shadow-mode results have been reviewed.
• Every exception or conflict. When documents disagree, or a case does not fit a defined outcome, it goes to an employee with the evidence attached.
• Any non-default outcome. For example, a validity period other than the standard one for that route, with the reason recorded.
• Rejections. A rejection cites the approved rule it relies on, and an employee reviews it before it is sent.
• Changes to rules and thresholds. A threshold change is an approved, documented action with the date from which it applies. It is not a quiet code change.

What can run without a person

• Asking the resident for a specific missing document, using approved wording.
• Attaching a newly submitted document to the original request and running the checks again.
• Closing cases that have a documented explanation and fall within approved rules, where the municipality has agreed to this for that process.
• Gathering supporting information into one case view, so the employee does not have to search several systems.

Make human and system actions distinguishable

The automation should run under its own dedicated user, not under an employee’s account, so the record shows which actions a person took and which the system took. Records should be append-only, including proof that a resident was notified.

Fail toward a person

If an automated step fails or times out, the case should go to an employee rather than stall or retry silently. A workflow that stops without telling anyone is worse than one that hands the case back.

Test before you trust

Run the workflow in shadow mode first: it proposes decisions without changing the system of record, and each proposal is compared with what an employee actually decided. Agree in advance what result allows the next stage, such as no incorrect approvals in the review sample.

Limits

Regulatory duties, notice periods and retention rules differ by service. A workflow can help make sure they are met on time; it does not shorten them. Each municipality decides its own approval points, and legal review is part of that decision.