A general-purpose AI tool is not built to write official minutes. In our own testing, general-purpose chatbots consistently made things up and missed things, getting roughly 20 percent of a set of minutes right. Even 80 percent would not be good enough, because minutes are a public record and the errors land in motions, votes and names. Minutes need AI with guardrails: speakers confirmed by the clerk, uncertain spots flagged instead of guessed, every line checkable against the recording, and nothing official until the clerk approves it.
We tried it ourselves. We took real meeting recordings, ran them through the general-purpose AI tools most people already have open in a browser tab, and asked for minutes.
The results were consistent, and not in a good way. The tools invented things that never happened, put motions in the wrong member's mouth, and skipped things that did happen. Across our tests they got about 20 percent of the minutes right.
That number gets attention, but it is not really the point. Suppose a tool got 80 percent right. For a meeting recap in someone's inbox, that might be fine. For the official record of a public body, it is not close.
Why isn't 80 percent good enough for minutes?
Minutes are not notes. They are what residents, auditors, attorneys and sometimes judges rely on to know what the board actually did. Three things make "mostly right" fail for that job.
The wrong 20 percent doesn't come labeled. An 80 percent draft has no map of its own mistakes. To find them, the clerk has to reread the whole thing against the recording, which is most of the work the tool was supposed to save.
The hard parts are the parts that matter. A summary of the discussion can be a little loose. Who made the motion, who seconded it, what it said and how each member voted cannot. Many states require the vote of each member, not just the result; Iowa, for example, requires minutes to show "information sufficient to indicate the vote of each member present."
Wrong minutes read just like right ones. A made-up sentence from an AI tool is grammatical, confident and plausible. Nothing about it looks wrong, so nothing prompts anyone to check it.
What does an AI mistake look like in minutes?
Here is the kind of error we mean. This is an illustrative example, not taken from a real meeting:
What happened: Member Lopez moves to table the paving contract until next month. Member Reed seconds. The motion fails, 2 to 3, with Members Lopez and Reed voting yes.
What a general-purpose tool might write: Member Reed moved to approve the paving contract. The motion carried unanimously.
Two sentences, three wrong facts: the mover, the motion and the result. If that goes into approved minutes, the official record says the town approved a contract it did not approve. We wrote more about why this happens in AI-generated meeting minutes and the public record.
Can you start your minutes in ChatGPT, Copilot or Gemini?
Yes, as a starting point. Paste in a transcript, ask for minutes, and ChatGPT, Copilot or Gemini will hand back something that looks like minutes in seconds. For a first pass at the discussion summary, that can save some typing.
The trouble starts when you try to turn that draft into the official record. This is where you are likely to get stuck:
- It doesn't know who sits on your board. Names come from context, so a motion can land on the wrong member, and nothing in the draft tells you.
- It won't tell you what it guessed. When the audio or transcript is unclear, it fills the gap with something plausible, and the guess reads exactly like a fact.
- It can't show you where a line came from. To check a motion or a vote, you go back to the recording and hunt for the moment yourself.
So you end up rewatching the meeting to check the draft, which is the work you were trying to save. These are not bad tools. They are built to write fluent, plausible text about anything, and they do that well. They are just not built for a public record. Accuracy in minutes does not come from a smarter model alone. It comes from the system built around the model.
What guardrails should an AI minutes tool have?
A guardrail is a rule that keeps an unchecked guess out of the record. These are the ones that matter for minutes:
- Speakers are confirmed, not guessed. The clerk confirms who each board member is, and the tool remembers them, instead of guessing names from context.
- Votes land on a named member. Movers, seconders and roll-call votes are tied to the person who spoke, not inferred from the outcome.
- Unsure spots are flagged, not filled in. When the tool is not confident, it says so and shows the clerk where to look.
- Every line traces back to the recording. The clerk can check any sentence against the moment it came from.
- A person approves before anything is official. The tool drafts. The clerk decides.
Those five do not make the AI perfect. They make its mistakes visible, which is what a public record needs. It is the same idea behind exception-based review: the tool handles what it is sure of and brings the clerk only what it is not.
What is the clerk's part of the guardrails?
The software is half of the system. The clerk is the other half. A checklist you can copy:
In the room
- Ask the chair to announce every vote result aloud, with nays and abstentions by name.
- Tally hand votes, arrivals and departures on your copy of the agenda.
On the draft
- Resolve every spot the tool flagged.
- Read each motion and vote against the recording.
- Check names, parcel numbers and dollar amounts against the packet.
- Read once for anything that should not be in the record.
- Approve only what you would sign.
Never: paste a recording or transcript into a general-purpose chatbot and publish what comes back.
Where Govably fits
Govably is built with these guardrails. You paste a link to the meeting recording. You confirm each board member once, and VoicePrint names them at every meeting after that. Movers, seconders and roll-call votes land on the right member. Unsure spots are flagged, not guessed, and Govably shows you exactly where to look. Click any line and the video jumps to that moment. You edit in the document editor, in your house style, and nothing is official until you approve it.
Our rule for building it is simple: we let actual intelligence check the artificial intelligence. The AI does the typing. The guardrails, and the clerk, keep the record yours.
Frequently asked questions
Can I use a general-purpose chatbot like ChatGPT to write our minutes?
You can use ChatGPT, Copilot or Gemini to start a rough draft, but not to write the official record. General-purpose chatbots are built to produce fluent text about anything, so when the audio is unclear they fill the gap with a plausible guess. In our own testing they consistently made things up and missed things, getting about 20 percent of a set of minutes right. Minutes need a tool that confirms speakers, flags what it is unsure of, and ties every line back to the recording.
Why isn't 80 percent accuracy good enough for meeting minutes?
Because minutes are the public record, and the remaining 20 percent does not come labeled. To find the wrong parts, the clerk has to reread the whole draft against the recording, which is most of the work the tool was supposed to save. The errors also tend to sit in the parts that carry legal weight: who moved, who seconded, what the motion said and how each member voted.
What is a guardrail in AI minutes software?
A guardrail is a rule built into the tool that keeps it from putting an unchecked guess into the record. Examples: speakers are confirmed by the clerk instead of guessed, votes are tied to a named member, uncertain spots are flagged rather than filled in, every line can be checked against its moment in the recording, and nothing becomes official until a person approves it.
Do we still need a clerk if we use AI minutes software?
Yes. Good software removes most of the typing and the rewatching, which is where most of the hours go. It does not remove the judgment. Someone still has to resolve what the tool flagged, confirm that each motion and vote matches what the board actually did, decide what belongs in the record, and stand behind the final document. That is the clerk's job, and the guardrails give the clerk the time to do it well.
What should an AI minutes tool do when it isn't sure?
Say so. A tool that guesses produces minutes that read smoothly and are wrong in exactly the places nobody checks. The better design marks the uncertain spot and points the clerk to that moment in the recording. A flagged spot takes a minute to resolve. A confident guess can sit in the approved record for years before anyone notices.
Source: Iowa Code § 21.3(2). This post is practical guidance, not legal advice; check your own state's open meetings law and your attorney for how it applies to your body.
Ryan Wilson is the founder and CEO of Govably, the AI minutes platform for local government clerks: cities, towns, counties and school boards.