What should staff never put into a public AI tool?

Anything you wouldn’t pin to a public noticeboard. That means residents’ personal details, HR files, legal advice, anything under a records hold, and decisions that aren’t public yet. The harder part is that staff already paste this kind of material into AI every day, so a list of rules is only half the answer.

Staff paste their own documents into AI every day

When local government staff use AI at work, they rarely start from a blank box. In GovAI, half of all messages start with someone pasting in their own document. Nearly half of messages include sensitive information of some kind: a name, an address, a case detail, an employee record.

That’s how the work gets done. The useful request is “summarize this complaint and draft a reply,” and the complaint has a resident’s name and address in it. A rule that says “never paste resident details” runs into the reason staff opened the tool in the first place.

Source: GovAI messages from US and Canadian local governments, to August 2026.

The front counter handles sensitive information too

Most people would guess that finance or HR puts the most sensitive material into AI. In GovAI, front counter requests carry more sensitive material than Finance’s. Labor relations, EMS, and social services carry the most.

That’s a useful correction. The staff answering residents’ questions at the counter, on the phone, and by email handle names, addresses, account numbers, and personal situations all day. They’re also the staff least likely to get specialist training on privacy. If your AI rules only cover the departments you think of as sensitive, they miss the counter.

Why local government has more to lose

A private company that leaks a customer email has a bad week. A local government has three more problems:

• Public records law. What staff write about city or county business may be a public record. If it was written in a personal AI account, nobody may be able to find it when a request comes in.
• Resident trust. Residents don’t choose their local government. They expect their details to stay with the people who need them.
• Public scrutiny. A council member, a reporter, or a resident can ask how AI is being used, and the organization should be able to answer.

A resident email, handled well

A front counter clerk gets an email from a resident asking why her permit is delayed. The thread includes her name, her address, the permit number, and two earlier replies from planning.

1. Strip what the task doesn’t need. To draft a reply, AI needs the question and the status. It doesn’t need her address or phone number. The clerk replaces them with [ADDRESS] and [PHONE].
2. Summarize instead of pasting everything. The two earlier replies can be one line: “Planning asked for a revised site plan on March 3. It hasn’t arrived.”
3. Ask for the draft. “Draft a polite reply explaining the permit is waiting on a revised site plan, and tell her how to send it.”
4. Check before sending. The clerk confirms the facts against the permit file, puts the details back, and sends it from her own account.

That takes a minute longer than pasting the whole thread. It’s a habit worth building, whatever tool staff use.

A simple rule for any AI tool

Share only what the task needs, and use the same judgment you would in any professional setting.

• Remove names, addresses, and case numbers before you paste a document in.
• Summarize the situation instead of pasting the whole file.
• Keep confidential or legal material out unless your organization has approved a tool for it.
• Take extra care with public communications, policy or legal content, and high-impact decisions. AI can still help. A person checks the result before it goes anywhere.

An approved place changes where staff paste

Many organizations respond to AI with a ban or a long list of don’ts. Staff who need to get the work done keep using AI anyway, on personal accounts IT can’t see. That’s shadow AI A ban alone doesn’t bring it back into view.

The safer move is to give staff an approved place to do the work, with clear rules and someone who can see how it’s used. In GovAI, your content stays yours, and it isn’t used to train AI models. Sensitive information is flagged as staff type, so the habit above gets a second check.

Check this week: where does your staff use AI today?

Five questions to ask your department leads. You don’t need a survey tool. A short conversation will do.

1. Which AI tools do people on your team use for work, including on their phones?
2. Are any of them personal accounts?
3. What kinds of documents go into them?
4. Does anyone on your team handle resident details at the counter, on the phone, or by email?
5. If a records request asked for everything written about a topic, could we find what was written in AI?

If the answers surprise you, the next question is about rules: Do we need an AI policy before staff start using AI?