AI prompts and outputs may be public records

If your city or county is starting to use an AI assistant, the clerk should ask one question first: is what it creates a public record? The honest answer today is that it may be, and you should decide before the first records request arrives, not after.

What gets created when staff use AI

Say a staff member asks an assistant to draft a memo from files on the city’s system. Three things now exist:

• The prompt, meaning what they asked
• The output, meaning what the assistant wrote
• The source files it drew on

Each of them could be a record. Whether they are depends on your state’s records law, your definitions, and what the content is used for. In Colorado, that starts with the Colorado Open Records Act, and your attorney is the right person to say how it applies.

This is not legal advice. It is the set of questions I would put in front of a clerk and a city or county attorney together.

The questions to ask

1. Where are prompts and outputs stored, and by whom? With an enterprise tool set up by the city, they sit in the city’s own systems. With a free consumer chatbot, they sit on a vendor’s servers, outside your control.
2. Does our retention schedule cover them? Most schedules were written before AI existed. They may not name this material at all.
3. Can a records request reach them, and who would search? If nobody knows how, the first request will be slow and expensive.
4. Who decides which prompts are records? A drafting request for a public notice is different from a private note to oneself.
5. What must never be pasted into a tool? Confidential legal, personnel, and public-safety items need a hard line.

What Microsoft tools can do

Cities on Microsoft 365 have some help. Microsoft says prompts and responses in Microsoft 365 Copilot can be covered by retention policies in Microsoft Purview. They are stored in a hidden folder in the user’s mailbox, and compliance staff can search them with eDiscovery tools.

That means the technology can support your answer. It does not choose the answer. Someone still has to decide what your schedule says, turn the settings on, and write down how a search works.

The consumer tool problem

The bigger risk is not the tool the city bought. It is the free one staff opened on their own.

If a department drafts public notices in a free chatbot, the city may have no copy of the prompt and no way to search for it. A draft council packet item pasted into a public tool has left the city’s control before it became public.

This is why a clear policy and short training matter as much as any setting.

Steps for a small team

You do not need a big project. Here is a workable order:

1. Find out what is in use. Ask each department, without blame.
2. Set the rule. Approved tools, and a list of what never goes in.
3. Update the retention schedule with your attorney to say what AI prompts and outputs are, and how long you keep them.
4. Turn on the settings in the approved tool so what the schedule says actually happens.
5. Write a one-page search procedure so a request has a known path.
6. Train staff, department by department.

Where this shows up in the assessment

It is sub-score 1.3, data quality and governance, and it connects to 1.4 and 3.3. In our AI Readiness Assessment for local governments, we present the findings at a public work session and note where records and retention rules touch each one.

If you serve on a council or run a clerk’s office and want to see how your city would grade, read how the assessment works.

Sources

• Learn about retention for Copilot and AI apps (Microsoft Learn)
• Search for and delete AI application data in eDiscovery (Microsoft Learn)