AI Meeting Notes: What the Transcription Gets Right, What the Summary Misses, and the One Line You Say Yourself

I sit in a lot of status meetings, and the only line I need from the notes afterwards is the decision. That is the line the AI summary most reliably misses, while it faithfully records the five minutes about the weather.

The transcription is good now, around 95 percent on clean English audio. The summary is readable. The action items arrive without owners. This post covers which kind of tool to use, how to get a transcript that is worth keeping, why the decision has to be said out loud in one sentence, and what to settle about consent before the bot joins the call.

Two steps, and they fail differently

Every meeting tool does the same two things.

Speech to text. A recognition model (OpenAI's Whisper, Google's Chirp, Meta's SeamlessM4T, or a vendor's own) turns the audio into words with timestamps. This step depends on the audio: microphone, background noise, accents, people talking over each other. It knows nothing about your project.

Text to notes. A language model reads the transcript, adds punctuation, labels the speakers, and writes the summary and the action items. This step depends on how well the model follows the thread of the conversation and how much of your jargon it recognises. It knows nothing about the audio.

When the notes are wrong, work out which step failed. A wrong product name is step one, and the fix is a better microphone or a custom vocabulary. A missing decision is step two, and the fix is saying the decision more clearly, or reading the transcript yourself.

When the vendor's summary is not good enough, paste the part of the transcript that matters into the AI text summarizer for a second, shorter pass. That tool sends the text to a server for the model call, unlike most of ToolForte, so give it the part you would put in an email anyway, not the confidential half hour. Then run the result through the readability checker before it goes to people who were not in the room. Notes written for the attendees are usually unreadable to anyone else.

A bot that joins the call, or a tool that takes a file

The meeting bots (Otter.ai, Fireflies.ai, Fathom, tl;dv) connect to Teams, Zoom or Meet, join the call as a participant, record, and produce the notes as the meeting ends. Their value is the plumbing: notes into Slack, tasks into Jira or Asana, a recording you can search by keyword. Teams and Google Meet now have their own version built in, on the paid tiers.

The transcription engines (Whisper, AssemblyAI, Deepgram) take an audio file and return a transcript. They work on anything: a phone recording of a meeting in a room, an interview, a voicemail. Turning the transcript into notes is your job afterwards.

For the recurring video meetings, use a bot. Ten to fifteen minutes saved per meeting, and the notes land where people look. For the quarterly session in a real room with a microphone on the table, record it, run the file through an engine, then summarise.

Test with your own audio, not the demo. A tool that scores 95 percent on a podcast can drop to 80 on four people sharing a laptop microphone. Record one ordinary meeting, run it through two candidates, and count the errors in the names and the numbers. That is the whole evaluation.

Better audio beats a better model

Nothing in step two fixes a bad step one, and the improvements cost nothing.

• One microphone per person, or one good one on the table. A laptop microphone records the keyboard and the room. A cheap USB microphone in the middle of the table is the biggest single improvement for in-person meetings. Remote, ask for headsets.
• One speaker at a time. The models handle crosstalk badly. A meeting norm of letting people finish is worth more than any setting.
• Feed it the names. Product names, internal acronyms, the customer's name. Most tools accept a custom vocabulary list. Fill it with the twenty terms that come up every week.
• Introductions, briefly. If the tool labels speakers, a round of "this is Toine, I have the test status" gives it a labelled sample of each voice.
• Two minutes afterwards. Read the transcript around the decisions and the numbers, and correct them. A misheard name in the transcript becomes a wrong owner in the action list.

The word counter on a transcript is a rough gauge of the meeting itself. Thirty minutes at 4,500 words was a discussion. Thirty minutes at 800 words was a presentation, and the meeting cost calculator will tell you what it cost to have everyone watch it.

Action items need a name and a date, said out loud

An action item has three parts: what, who, and by when. The tools get the what. They get the who and the when only if someone said them.

"Someone should look at the billing issue" becomes an action item with the owner "unassigned". "Sarah will look at the billing discrepancy by Friday" becomes an action item you can put in a tracker. The difference is not the tool. It is the sentence spoken in the meeting.

The same goes for the decision. If it was reached over four minutes of back and forth, the summary will paraphrase the back and forth. Say it once, plainly, at the end: "So the decision is to move the release by two weeks and fix the login first." One sentence like that is what the model extracts, and it is what the people who were not there need.

After the meeting, five minutes on the action list: add the missing owners and dates, delete the items that were musings, and send it. The five minutes you spend on the list save everyone else the hour of trying to remember what they agreed to.

Consent, storage and the meetings that should not be recorded

A recording of a meeting contains strategy, figures, names and sometimes a conversation about a person. Before a bot joins every call in the company, settle four things.

Consent. In the EU, GDPR applies to the recording and the transcript as personal data. In the Netherlands you may record a conversation you take part in, and you still owe the other participants openness about what happens to it. In the US it depends on the state, and some require everyone's agreement. The bot's join notification is a start. A line in the invite and a sentence at the top of the meeting is better.

Storage. Where the recordings live, for how long, and who can search them. If your meetings touch regulated data (health, finance, government), check the vendor's certifications against your own obligations before the pilot, not after.

Training. Some vendors use customer audio to improve their models and some do not. For meetings with anything confidential in them, pick one that does not, or run Whisper on your own hardware.

Retention. Set automatic deletion. A stand-up transcript from three years ago has no value and a real liability.

And some meetings do not get recorded at all: performance reviews, anything legal, anything about a person's health or contract. Make it easy to ask for the bot to leave, and do not argue when someone asks.

The workflow around the tool

The notes are only useful if someone reads them within a day.

• Before: share an agenda. Some tools use it to structure the summary, and it keeps the meeting on the topics you will want notes about.
• During: let the bot run and forget it. Restate each decision in one sentence when it is reached.
• Within half an hour after: read the summary, fix the errors, add owners and dates.
• Same day: send the notes where people already look. A wiki page nobody opens is the same as no notes.
• Two days later: one message to each owner. Did you see it?
• Friday: all the open items from the week in one list. The AI text summarizer can compress a week of notes into the paragraph a steering committee will actually read.

FAQ

How accurate is transcription with several speakers?

With headsets or a good table microphone, 90 to 95 percent. Four people on one laptop in a noisy room, 70 to 80. Speaker labelling is less reliable than the words, and worst for voices that sound alike.

What about meetings in two languages?

Most tools want one language per meeting. A Dutch meeting that switches to English for the one colleague who does not speak Dutch, and back, still confuses most of them. The multilingual models cope better and are not yet the default inside the meeting bots.

Can the transcript serve as a legal record?

Not as it comes out of the tool. Recognition errors, missing context and mislabelled speakers make raw output unreliable for anything with consequences. A person reviews and signs off any transcript used for compliance or a dispute.

What does it cost?

The meeting bots run from free with limits to about 15 to 30 dollars per user per month. The engines charge per minute of audio, roughly half a cent to two and a half cents. Whisper on your own machine is free and needs a decent GPU or patience.

Do I need everyone's permission to record?

In the EU, tell everyone, have a legitimate reason, and treat the recording as personal data. In parts of the US, everyone has to agree. Write the rule down once, put it in the invite template, and the question stops coming up.