Can AI write meeting minutes? An honest look for clerks

AI tools that turn a meeting recording into draft minutes are now common, and plenty of clerks are being asked, "Can't the computer just do that?" The honest answer is: it can do a useful first draft, and it cannot do the job. This article covers where the tools help, where they fail, and what to ask before you let one near your board's records.

How these tools work

Most AI minutes tools do two separate things:

Transcription. A speech-recognition model turns the audio into text, often with timestamps and an attempt to separate speakers.
Drafting. A language model reads the transcript and writes minutes: attendance, motions, votes, and summaries, in the format you ask for.

Errors can come from either step, and they compound. A misheard name in the transcript becomes a wrong mover in the minutes.

What AI does well

Getting the words down. With clean audio, modern transcription is fast and generally accurate for ordinary speech. It saves the most tedious part of the job: replaying the recording to find what was said.
Structure. Given an agenda, a drafting model is good at putting discussion under the right item and producing consistent formatting.
Finding the motions. Phrases like "I'll move that," "second," and "all in favor" are distinctive, and models are good at spotting them.
Summaries. For summary-style minutes, a model can condense a 20-minute discussion into a few neutral sentences that a clerk can then edit.
Search. A timestamped transcript makes it easy to jump to the moment a vote happened and check it.

What AI does badly

Speaker attribution

Knowing who said something is harder than knowing what was said. Speaker separation ("diarization") often confuses people with similar voices, people who sit far from a microphone, and people who talk briefly. Even when it separates voices correctly, the tool still has to work out that "Speaker 3" is Supervisor Chen. For minutes, attribution is exactly what matters: who moved, who seconded, who voted no.

Vote counts

Voice votes are a particular problem. "All in favor?" followed by a murmur of ayes doesn't tell the software who voted, and a quiet "no" can be missed entirely. Roll calls are easier, but a tool may still attach a vote to the wrong name. Never trust a vote count you haven't checked against the recording or your own notes.

Crosstalk and off-mic speech

Two people talking at once, a member leaning away from the microphone, or a comment from the audience can produce garbled or missing text. The drafting step may then smooth over the gap and produce confident-sounding text that isn't what was said.

Hallucinations, including during silence

Speech-recognition models can produce text that was never spoken. Researchers studying OpenAI's Whisper found that roughly 1% of audio transcriptions in their study contained entire hallucinated phrases or sentences, and that hallucinations occurred disproportionately for speakers with longer pauses and non-vocal stretches (Koenecke et al., "Careless Whisper"). Public meetings have plenty of silence: recesses, members reading documents, waiting for a slide. Treat any text that appears during a quiet stretch with suspicion.

Context the recording doesn't contain

A recording doesn't show that a member left the room during an item, that a document was handed out, or that the chair nodded to end a discussion. It doesn't know your board's usual practices or which items were on the consent agenda unless you tell it.

Why human review is non-negotiable

Minutes are an official public record, approved by the board and relied on for years. The clerk or secretary is responsible for their accuracy, and the board is responsible for approving them. AI doesn't change who is accountable.

That means:

AI output is a draft, never the minutes.
A person who attended the meeting, or who has checked the recording, reviews every motion and vote.
Nothing is posted as minutes until the clerk has finished review, and nothing is "approved" until the board approves it under its normal process.

A good tool makes this easier by flagging what it isn't sure of, rather than hiding uncertainty behind clean formatting.

Questions to ask vendors

Your recordings and transcripts may include executive-session audio, residents' names and addresses, and personnel matters. Before signing up, get written answers to these:

Training and use of data

Do you or your AI providers use our audio, transcripts, or minutes to train models?
If there's an opt-out, is it the default?
Which third-party AI services process our data?

Retention and deletion

Where is the audio stored, and for how long?
Can we delete recordings and transcripts, and does deletion include backups and subprocessors?
What happens to our data if we cancel?

Location and security

Where is the data processed and stored geographically?
Is it encrypted in transit and at rest?
Who at the vendor can access it?

Public records

Can we export everything (audio, transcript, drafts, final minutes) in standard formats?
Does the tool keep draft versions, and can we control that?

The last question matters because recordings, transcripts, and even drafts may be public records under your state's law, with their own retention rules. Some states set specific rules for particular recordings. Illinois, for example, requires verbatim recordings of closed sessions and sets conditions for when they may be destroyed (5 ILCS 120/2.06). Ask your records officer or state archives how your retention schedule treats recordings and transcripts before you adopt a tool. Also check whether closed-session audio should go through any outside service at all.

This is general information, not legal advice.

A review checklist for AI-drafted minutes

Work through the draft with the recording open.

Header and attendance

Meeting type, date, start time, and location are correct
Every member present and absent is listed correctly, including remote attendance
Late arrivals and early departures are recorded with times
Quorum statement is accurate

Motions and votes

Every motion on the recording appears, and none were invented
Mover and seconder are correct for each
Motion wording matches what was voted on, including amendments
Each vote result and each member's vote is correct
Abstentions and recusals are recorded as they happened

Content

Dollar amounts, names, addresses, and parcel or resolution numbers are exact
No text appears during silences or recesses that wasn't spoken
Summaries are neutral and don't attribute opinions to the wrong person
Nothing from closed session appears in open-session minutes
Public comment follows your state's rules and board policy

Finish

Every flagged or uncertain item has been resolved
The document is marked "Draft" until the board approves it

If the tool is saving you time on transcription and structure, it's doing its job. If you find yourself trusting it on votes, slow down.

Where CivicPorch fits

CivicPorch drafts minutes from a recording, marks anything it can't confirm with "⚠ check" for you to resolve, and never publishes minutes without clerk approval. It's free for one board, including two AI-drafted meetings, if you'd like to try it.