Thereâs a thing every meeting-note AI tool advertises that nobody ever talks about honestly: a transcript is not minutes. A transcript is 14,000 words of âyeah, totallyâ and âsorry, you cut outâ. Minutes are five decisions, three action items, and one risk. Getting from one to the other is the work, and the tool you pick determines how much of that work is left for you.
Most assistants pick a tool based on a free trial and a headline feature. Better to pick based on what your meetings actually are. Hereâs how the four most common options actually compare in admin work, plus the prompts that turn whatever the tool gives you into something your exec will read.
The four tools, what each is good at
Otter.ai is built around a clean web app and a phone app. Strong at speaker labelling once itâs heard each voice a few times. Free tier covers 300 minutes a month, paid plans go up from there. Best fit when youâre sitting in a room with a laptop on the table and a meeting that isnât on a video platform.
Microsoft Copilot inside Teams writes the recap natively after a recorded meeting. The summary it generates is already structured into headings (decisions, action items, open questions). Best fit if your org runs Teams and you want one fewer tool to manage. The catch is it only sees what got recorded, and only if recording was on.
Google Gemini inside Google Meet works similarly: âTake notes for meâ runs through Gemini and produces a recap doc in Drive. Best fit for Google Workspace shops. Same recording-required catch.
Zoom AI Companion works the same way for Zoom calls. Itâs free at the basic tier inside paid Zoom accounts, which is the lowest-friction option if your execâs external meetings happen on Zoom.
The real comparison isnât feature lists, itâs the meeting context. If your exec runs every meeting in the same platform, use that platformâs native AI and skip the third-party tool. If meetings sprawl across Zoom, Teams, and in-person, Otter is the one that travels with you.
The consent rule
Before any of this matters: in many jurisdictions, recording a meeting (or putting an AI bot in it) requires the consent of every participant. New York, California, Florida, Pennsylvania, Washington and others are all-party consent states. The UK and EU treat recording without consent as a data-protection issue. Australia varies by state.
The rule that keeps you out of trouble: at the start of every meeting where you have AI transcription on, the host says aloud, âIâve got [TOOL] taking notes today, the transcript will be shared with [WHO], let me know if youâd prefer I turn it off.â Then put that line in your meeting templates so nobody forgets. One line, every meeting, no exceptions. The day you skip it is the day someone says something they wouldnât have said if theyâd known.
The prompt that makes minutes from a transcript
Whatever tool you used, you now have a transcript or a raw recap. Open ChatGPT or Claude, paste it in, and run:
âYouâre producing meeting minutes for [EXEC NAME] from this transcript. The audience is [INTERNAL TEAM / EXTERNAL CLIENT / BOARD]. Format:
Attendees: [LIST FROM TRANSCRIPT, MARK ANYONE ABSENT] Date and length: [FROM TRANSCRIPT]
Decisions made (each one to two sentences, only if a real decision was reached, not âwe discussedâ).
Action items (table: owner, action, deadline, source quote in italics so I can verify).
Open questions (anything raised that wasnât resolved, with who needs to answer it).
Risks or concerns flagged (only if explicitly raised, donât editorialise).
Attach a one-paragraph summary at the very top, under 80 words, written so my exec can skim it in 15 seconds and know what happened.
Donât include small talk, status updates that already had a decision, or repetitions. If a section has nothing in it, write âNone recordedâ rather than padding.â
The âsource quote in italicsâ line is the load-bearing one. It gives you a way to verify each action item without re-reading the whole transcript, and it surfaces the cases where the model has interpreted âIâll think about itâ as a commitment.
A worked example, before and after
Suppose a 47-minute project status call. The raw Otter transcript is 9,200 words. The default âsummarise thisâ prompt gives you a paragraph that reads like a press release: âThe team had a productive discussion about the Q3 launch and aligned on next steps.â Useless.
Run the prompt above and you get something like:
âSummary: The Q3 launch ships 14 October. Marketing missed the asset deadline and is now blocking design review. Sarah will escalate to her VP today; if not unblocked by Friday, the launch slips a week.
Decisions: Launch date confirmed 14 October. If marketing assets arenât approved by Friday EOD, the team triggers the contingency plan and slips to 21 October.
Action items: Sarah to escalate to marketing VP by EOD today. Tom to draft the contingency comms by Wednesday. Priya to confirm legal review on Thursday.
Open questions: Who owns the post-launch metrics review? Raised by Sarah, no answer in this meeting.â
Same transcript, completely different artefact. The exec reads the summary in ten seconds, opens the action items if anything looks off, and moves on. Thatâs the standard.
The failure mode that gets minutes thrown out
Confabulation. Models will sometimes fabricate an action item or a decision that nobody actually made, especially in long, rambling conversations. Two protections.
First, the source-quote requirement above. If a model canât produce a verbatim quote backing an action item, that item probably isnât real. Pull it.
Second, send the draft minutes to one attendee for a sanity check before you publish to the whole list. Pick the person who said the most in the meeting, since theyâre most likely to spot whatâs missing or invented. Two-minute scan, catches 90% of errors. Skipping this step is how an action item that nobody agreed to ends up on a status report.
A counterintuitive note on quotes
The cleanest minutes donât include quotes. Direct quotes feel transparent but they make minutes longer, harder to scan, and politically risky (someone always wants their phrasing softened after the fact). Use quotes in your verification step. Strip them from the version that goes out.
What the tool canât do
The tool transcribes. The model summarises. You pick what gets recorded, who gets the minutes, whatâs omitted because it was an off-the-record aside, and which action item needs to be flagged to your exec personally because itâs politically sensitive. Those are still judgement calls. The work thatâs left after AI does the typing is the work that mattered all along, just compressed into 15 minutes instead of 90.
The administrative assistants writing the cleanest minutes now arenât transcribing faster, theyâre choosing better. The transcription is automated. The choosing is the job.