How can a small municipality use AI responsibly?

A town of a few thousand people may have a clerk, a treasurer, a public works lead and a handful of others doing everything from permits to the council agenda. They write constantly: notices, summaries, letters, minutes. Generative AI, the kind of tool that produces text from a short instruction called a prompt, can take some of that load. It can also, used carelessly, publish a wrong date, leak a resident's information or erode trust in a public institution.

The good news is that a small municipality doesn't have to invent its own rules. Canada's federal government has published detailed guidance for its own employees, privacy commissioners have set out principles for any organization using these tools, and in Québec, Law 25 already spells out duties for public bodies. Together they give a sensible starting framework.

Start with low-risk uses

The Government of Canada's Guide on the use of generative AI, updated in July 2024, recommends that institutions experiment with low-risk uses before higher-risk ones. Its examples of low risk include an internal email and editing a draft that will be reviewed anyway. Its examples of higher risk include a chatbot for the public and summaries of client information.

For a town hall, that ordering looks like this:

Worked example: a bylaw summary and a bilingual notice

A new bylaw on outdoor watering is adopted. The full text is public, so it can be pasted into an approved tool.

Below is the full text of a municipal bylaw adopted by our council.
Write a summary for residents of no more than 150 words, at a grade 8 reading level.
Include: who it applies to, what is allowed and when, the start date, and how to ask questions.
Use only information in the bylaw. If something residents would need is missing, list it separately as "Not in the bylaw."
Then write the same summary in English.

[paste the bylaw text here]

The "Not in the bylaw" line matters: it stops the model from filling gaps with plausible inventions, such as a fine amount that isn't in the text. Before publishing, a staff member compares every date, hour and amount with the bylaw, and a fluent speaker reviews the translation. The federal guide warns that these tools can perform unevenly from one language to another, so a second-language version is always a draft until a person approves it.

Then label it. The guide's transparency principle asks institutions to identify content produced with generative AI. A line such as "Summary prepared with the help of an AI tool and verified by municipal staff; the bylaw text prevails" is honest and reassuring.

The rules that make it responsible

Keep personal information out of public tools. The federal guide is direct: public servants must not enter personal information into publicly available generative AI tools, because the supplier may keep a copy. The privacy commissioners' principles add that de-identified information should be used in prompts wherever possible. A complaint letter, a tax account or a permit file stays out of any consumer chatbot.

Follow the FASTER principles. The federal guide sums up responsible use in six words: Fair, Accountable, Secure, Transparent, Educated and Relevant. In practice: check outputs for bias, keep a person responsible for every published text, use tools suited to the sensitivity of the information, disclose AI use, train staff, and use AI only where it actually helps. The guide is candid that, for some jobs, AI is simply not the right tool.

Meet your Law 25 duties. Law 25 amended Québec's access and privacy act for public bodies, and most of its provisions took effect on September 22, 2023. Among the new obligations, public bodies must adopt rules governing how they handle personal information, and must carry out a privacy impact assessment (a structured review of privacy risks) for any project to acquire, develop or overhaul an information system or electronic service that involves personal information. A resident-facing chatbot, or an AI feature added to your permit software, falls squarely in that category. Talk to your access-to-information and privacy officer before signing anything.

Write it down. A one-page policy adopted by council (approved tools, forbidden data, review steps, disclosure wording) protects staff as much as residents.

Which tasks should stay with staff?

Anything that decides something about a person. Permit approvals, fines, tax adjustments, eligibility for programs and responses to complaints carry legal and human consequences, and residents rightly expect a decision from someone accountable who can explain it. AI can help a staff member find the relevant bylaw article; it shouldn't produce the decision.

Sensitive conversations also stay human. A resident in distress, a dispute between neighbors, a call about an emergency: these need judgment, empathy and local knowledge a model doesn't have. The federal guide asks public servants to form their own views before asking an AI for recommendations and to use these tools as aids rather than substitutes, which is good advice for a council briefing note, too.

Finally, the last check before anything goes public belongs to staff. A model can draft a notice in seconds, but the person who signs it answers for every date and every word. In a small community, where residents know the people at the town hall by name, that personal accountability is part of the service.