AI meeting minutes start as a draft, not an approved record. A transcript can help structure discussion, decisions and action items. It cannot independently verify facts that were never spoken: who was absent, how a member voted by gesture, or which deck was displayed without being named.

That gap does not look like an error. A transcript with a missing sentence may be obviously incomplete. Draft minutes that omit an unspoken abstention can still look finished.

Three US open-meetings statutes illustrate why a public body's minutes need information beyond a transcript. This guide focuses on what the audio can establish. It offers a prompt that marks missing facts, checks the statutory examples, separates word accuracy from speaker attribution, and examines multilingual summaries.

Key Takeaways

A prompt for AI meeting minutes that flags its own gaps

Most minute-drafting prompts ask for a format. Also tell the model what to do when it does not know. A plausible-looking draft can hide missing facts; make uncertainty visible for the person who reviews it.

Prompt — copy and paste
You are drafting minutes from a meeting transcript. Work only from the
transcript. Do not infer, and do not fill gaps from convention.

Produce these sections:
1. Meeting - date, time, place, name of the body.
2. Attendance - members present, members absent, others present.
3. Items - for each agenda item: subject, a 2-4 sentence summary of the
   discussion, the exact wording of any motion, who moved it, who
   seconded it.
4. Votes - for each motion: the result, and the name of every member
   recorded as voting against or abstaining.
5. Documents - every exhibit, deck, report or handout referred to.
6. Actions - task, owner, due date.

Rules:
- Where the transcript does not contain a required fact, write
  "NOT IN AUDIO" on that line. Never guess it, and never drop the line.
- Copy a motion only if its wording is clear in the transcript;
  otherwise mark its exact wording NOT IN AUDIO. Paraphrase discussion.
- Label any voice you cannot attribute as "Unidentified speaker".
  Do not assign it to a named person.
- Finish with a section headed "Fill in before circulating" that lists
  every item marked NOT IN AUDIO.

Transcript:
<paste transcript here>

The format lines are ordinary. The value is in the gap rule. A draft that says Members absent: NOT IN AUDIO gives the secretary a visible item to verify. The list still needs a human review: a model can miss a fact that actually was spoken, or fail to flag an omission. If you are choosing a transcription tool, our comparison of AI meeting note takers covers capture choices.

Working across languages? MirrorCaption streams transcription and translation side by side while people are still speaking, so you can annotate the gaps during the meeting instead of after it.

Try it free

What minutes must contain that the audio never held

Public bodies must follow the law that applies to them. These three state examples show details that audio alone may not establish. Massachusetts law explicitly lists documents and exhibits. A private team's minute requirements will differ, but it can use the same verification method.

Required in the minutesSourceDoes it make a sound?
Names of the members present Georgia, O.C.G.A. § 50-14-1(e)(2)(A) Only for the ones who speak or are addressed by name
Names of members in attendance and those absent New Mexico, NMSA 1978 § 10-15-1(G) Not without a roll call or another spoken record
Who made and who seconded each motion Georgia, O.C.G.A. § 50-14-1(e)(2)(B) Usually, if the chair restates it aloud
The name of each member voting against or abstaining Georgia, O.C.G.A. § 50-14-1(e)(2)(B) Not for a raised hand or a headshake
A list of documents and other exhibits used Massachusetts, G.L. c.30A § 22(a) Not if the exhibit was never identified aloud

Georgia turns the missing line into a yes

The fourth row is the one worth sitting with. Georgia's published code sets a presumption for how the minutes are read when they omit the names of dissenting or abstaining members:

"It shall be presumed that the action taken was approved by each person in attendance unless the minutes reflect the name of the persons voting against the proposal or abstaining."

O.C.G.A. § 50-14-1(e)(2)(B)

If a member votes against a motion by a permitted nonverbal method, the audio may contain no trace of that vote. A transcript-based draft can then omit the member's name. Georgia's presumption makes the omission consequential for readers of the minutes; the clerk must check the actual vote record and correct the draft before approval. The statute does not turn an actual no vote into a yes vote.

Where an AI draft genuinely helps

The same statutes show why a timely draft helps. Georgia requires a summary of subjects acted on and members present to be available within two business days of adjournment. New Mexico's Open Meetings Act guide says draft minutes must be prepared within ten working days. A transcript can speed up the discussion summary, while the clerk supplies the missing attendance, votes and exhibits from other records.

Illustrative scenario

At a hypothetical five-member commission meeting, four members say "aye" and the fifth casts a permitted no vote by raising a hand. A transcript-based draft lists all five as present but does not name the dissenter. Under Georgia's presumption, a reader of those uncorrected minutes would infer that every attendee approved the action. The clerk checks the vote record and adds the fifth member's name before the minutes are approved.

Why "99% accurate" measures the wrong thing for minutes

Fireflies, for example, advertises 99% transcription accuracy in English and 95% in other languages. Its blog does not specify a test set or scoring method for those figures, so they should not be treated as a measured guarantee for a particular meeting.

Minutes also depend on speaker attribution: who said the words. A separate diarization error rate (DER) measures missed speech, false alarms and speaker confusion. pyannoteAI reports 17.0% DER for its Community-1 model on AMI close-talk microphones, 19.9% on AMI single distant microphone, and 20.2% on DIHARD 3. Those are specific benchmark conditions, with fully automatic processing, no forgiveness collar and overlapping speech included.

The Fireflies and pyannoteAI percentages use different systems and metrics, so comparing them as if they were competing accuracy scores would be misleading. Neither figure predicts the quality of minutes from a new tool or room. Check speaker labels against the audio and attendance list, especially before attributing a motion to a named person. Our transcription accuracy guide explains why benchmark conditions matter.

Evidence about summaries is not the same as evidence about statutory minutes. A 2026 evaluation of AI meeting summaries tested 114 meetings across three domains. Under its protocol, differences in accuracy among the tested models were not statistically significant after Holm correction (adjusted p-values 0.053–0.448); the paper did find significant separation in information retention, with GPT-5.1 leading on completeness (0.886) and coverage (0.942). Those scores do not measure whether a draft captured an unspoken vote or satisfied a state's minutes requirements.

What happens to AI meeting minutes in a bilingual room

Check the language behavior of the tool actually used. Fireflies says its Multi-Language Mode is in beta on Business and Enterprise plans, covers 60+ languages in a mixed-language session, and produces a summary in the dominant spoken language even when its transcript contains several. Other tools may behave differently.

For formal minutes, a translated paraphrase deserves review against the original speech. A concern voiced in Spanish during an otherwise English meeting could be shortened in the English summary; the secretary should check that the final minutes preserve its meaning.

Illustrative scenario

A hypothetical works council meets mainly in German with two English-speaking observers. The chair briefly switches to English, and a tool later writes its summary in German. The observers ask to review how their remarks were summarized before those words enter the draft minutes. The original transcript and translation let the secretary resolve any disputed phrase.

Separate live capture from drafting. Seeing the original transcript and translation side by side can help the secretary flag a disputed phrase during the meeting. MirrorCaption offers live transcription and translation in 60+ selectable languages without a meeting bot; it does not create or approve official minutes. Our guide to real-time voice translation for business calls covers translated speech output.

Five checks before you circulate AI meeting minutes

None of these require a better tool. They are the places where a human has information the audio does not.

  1. Check attendance against the invite list, not the transcript. The draft knows who spoke. Someone who joined, listened, and left is invisible to it — and in New Mexico, so is everyone who never joined at all.
  2. Confirm quorum at the moment of each vote, not just at the start of the meeting. People may drop off; verify the headcount against attendance and the meeting platform or clerk's record.
  3. Resolve every "Unidentified speaker" line before it reaches the record. The cited DER scores do not forecast your tool's error rate, but any attribution that matters should be checked.
  4. Reconstruct nonverbal votes from the official vote record. If the meeting rules permit gestures or electronic votes, verify the names of those voting against or abstaining before approval.
  5. List the documents. Anything shown on screen and never named aloud exists nowhere in the transcript. Pull it from the agenda or the share history.

Two of the five are about people who made no sound, and two are about the vote itself. That distribution is the whole argument of this page in miniature.

See both languages while the meeting is still running

Live transcription and translation in your browser. No bot joins the call. Eligible new users can try one free hour without a credit card.

Get Started Free

Frequently asked questions

Are AI meeting minutes accurate enough to be the official record?

The public body must review and approve its minutes under the rules that apply to it. A transcript-based draft may omit attendance, quorum or nonverbal votes unless someone states those facts aloud; check them against the meeting record before approval.

What is the best prompt for AI meeting minutes?

Use a prompt that asks the model to mark gaps rather than fill them. Request the standard sections, label required facts absent from the transcript as NOT IN AUDIO, and include a closing list of those labels. Review the draft for missed facts. The full version is above.

Can AI take minutes for a board meeting?

AI can draft a discussion summary and identify motions stated aloud, subject to checking the transcript. A secretary still needs to confirm quorum, absences, exact motion wording and nonverbal votes against other meeting records.

What happens if the meeting is in more than one language?

Language handling depends on the tool. Fireflies says its Multi-Language Mode can transcribe several languages while the summary uses the dominant spoken language. Review a translated summary against the original speech before adopting it as a record.

Do AI meeting minutes require recording the meeting?

Not necessarily. Retention depends on the tool and its settings. MirrorCaption can transcribe a live audio stream without saving a recording; its optional playback recording is stored locally on the device. Our explainer on what AI meeting tools do with your data compares policies.

The short version

A transcript can support a minutes draft, but it cannot independently establish unspoken facts. Confirm quorum, absences, exhibits and nonverbal votes from meeting records. Georgia's presumption shows why a missing name in the minutes matters even when the audio offers no clue.

Use the prompt above to mark gaps, then check attendance, quorum, speaker labels, votes and the document list before circulation. For multilingual meetings, compare any translated summary with the original speech. Open MirrorCaption in your browser to view live captions and translation — eligible new users can try the free hour without a credit card.