AI for Municipal Staff: Start With One Workflow You Can Measure

Choose a task with a clear starting point and finish

A clerk preparing a routine service notice may spend time turning the same approved facts into a website update, newsletter paragraph, and staff reference. AI could help prepare those drafts. The useful question is whether the complete process becomes easier after someone checks the dates, contacts, and instructions. A fast first draft is only one part of the work.

For Michigan townships, villages, and cities, choose one repeatable task whose result staff can readily verify. Examples include adapting a published notice, preparing a document summary, or organizing questions from an approved email thread. Name the employee who owns the workflow and the person who approves the finished work. Begin with one task before connecting more departments or information sources.

Give the draft a small, approved set of facts

For a hypothetical village facility closure notice, assemble the approved closure dates, affected location, alternative arrangements, and contact details. Ask the selected tool to use only that material and identify missing information. Specify the audience, length, and format. A staff member can then compare each version with the same source, rather than asking the tool to invent the office’s procedures.

An example instruction is: Prepare a short website notice from these approved facts. Preserve the dates and contact information exactly, and list any detail you need confirmed. Give the reviewer the original material alongside the draft. Check links, phone numbers, opening hours, and any suggested next step before using the municipality’s normal publishing process. The task owner remains responsible for the notice.

Use summaries to navigate the source material

AI can also help staff orient themselves within an approved document or discussion. Microsoft documents email-thread summaries in Copilot for Outlook and document summaries in Copilot for Word. If your organization has approved those capabilities, try a staff workflow such as identifying proposed next steps in a project email thread. Verify who proposed each action, its status, and any date against the original messages.

Outlook summaries may include citations to messages, but a citation still needs checking. Microsoft also notes that Word does not always provide citations for information later in a document. Keep the full source available and ask the responsible employee to check important details and omissions. An AI summary should not independently establish a municipal decision, change an approved policy, or become official meeting minutes without the established staff review and approval.

Agree on the tool and information before the pilot

Have IT and the workflow owner confirm the approved account, product features, licensing, and information allowed in the pilot. Start with public or appropriately approved sample material. Establish where prompts and outputs are stored, who can access them, and which integrations or web-search features are enabled. Sensitive resident, personnel, or other restricted information needs a specific approved handling process.

Microsoft’s Microsoft 365 Copilot documentation explains that organizational content is accessed through users’ existing permissions. Review the underlying sharing arrangements before using that content with an assistant; broad access does not become appropriate simply because an AI tool can use it. Different products and optional features can have different data handling. Ask IT to review the selected configuration rather than assuming one vendor’s protections apply to every AI account.

Give the reviewer a job they can complete

NIST’s Generative AI Profile identifies confidently incorrect output as a risk and recommends checking generated sources and citations. Build that into the task: compare factual statements with the approved documents, confirm that a cited passage supports the claim, and flag unsupported additions. A polished paragraph can still contain a wrong date or an instruction that does not appear in the source.

Define what happens when the result is incomplete or uncertain. The employee can revise it, return to the original workflow, or ask the task owner for clarification. Keep decisions about an individual resident, application, payment, or personnel matter with the authorized people and processes. Use the pilot to assist preparation of work, with a named person accountable for checking and approving the result.

Measure the complete job before expanding

Record a baseline for a few representative examples using the current method. During the pilot, record time spent preparing inputs, generating drafts, reviewing sources, correcting mistakes, and approving the final result. Note errors, missing information, and repeated instructions staff had to add. Compare the complete work against the baseline so checking and cleanup remain visible.

Agree on what would justify continuing: an accurate usable result, a manageable review burden, and a process employees understand. Include subscription and configuration costs in the review. NIST cautions against generalizing performance from narrow or anecdotal tests. Repeat the assessment with varied examples, and review it after changes to tools, source information, or staff responsibilities. Expand when the evidence supports the next workflow.

Use this checklist for the first staff workflow

IT for townships should connect new tools with the work people actually do. Bring one repeatable task to your next Michigan IT services conversation and agree on these points before starting.

• What task are we improving, and who owns the finished result?
• Which tool, account, features, and information are approved?
• Can the reviewer inspect the complete authoritative source?
• Which facts, omissions, and citations must be checked?
• How do we record preparation, review, correction, and total time?
• What results and costs will determine whether the pilot continues?