The bot that joins your meeting
Someone adds a notetaker to the invite. A new participant appears in the list with a name like Notetaker or Assistant, sits silently through the hour, and shortly afterwards a tidy summary arrives with headings and bullet points. It feels like magic, and on a good day it genuinely is.
It is also a machine with specific and fairly predictable blind spots. The people who come unstuck are not the ones who never used a notetaker. They are the ones who used one for a year without ever learning what it is bad at, and then forwarded a summary that said something untrue about a colleague.
In plain English
- AI notetaker:
- A tool that listens to a meeting, writes down what was said, and produces a summary afterwards.
- Transcript:
- The word for word record of what was said, with speaker labels. Long, messy, and almost nobody reads it.
- Summary:
- The short version the tool writes from the transcript. This is the bit people actually read.
- Speech to text:
- The step that turns sound into words. Where accents, mumbling and crosstalk cause most of the damage.
- Speaker labelling:
- Working out who said which bit. This is the step that puts the wrong name next to a sentence.
Three jobs pretending to be one
Almost every product in this category does the same three things in the same order, and knowing which stage failed is most of the skill of using one well.
It captures. Something listens, whether that is a bot sitting in the call, a feature built into the conferencing software, or an app running on a laptop in a room. At this point all it is doing is getting hold of audio.
It transcribes. The audio becomes text with speaker labels attached. This stage is not thinking about meaning at all. It is matching sound to words.
It summarises. A language model reads that text and writes the short version: themes, decisions, actions. This stage never heard the meeting. It only ever sees the transcript.
That last point is worth carrying around with you. If the transcription stage mangles a name, the summary confidently repeats the mangled name, because as far as the summariser is concerned that is simply what the person is called. Errors do not get caught further down the line. They get tidied up and made to look official.
Where each stage struggles
The capture stage struggles with rooms. A laptop at one end of a boardroom table hears the two people nearest it beautifully and the person at the far end as weather. It struggles with the colleague dialling in from a car, from a station platform, or from a kitchen with the dishwasher running.
The transcription stage struggles with anything unusual, and unusual here means unusual relative to the enormous quantity of everyday speech these systems learned from. Strong regional accents. Colleagues speaking English as a second or fourth language. Meetings where people switch between two languages in the same sentence, which is entirely normal in a great many workplaces and reliably confuses the machine. Your industry's jargon. Your product names. Acronyms your company invented last spring. And, most stubbornly, people's names, which is the error that causes genuine offence and which lesson five deals with properly.
Crosstalk is its own category. When three people talk over each other because the idea is exciting, a human in the room follows it without effort. The transcript turns into a jumble, or drops one voice entirely, and the summary quietly reports whatever survived as though it were the whole discussion. The person who was talked over does not appear to have spoken.
The summarising stage struggles with anything that is not literal. A joke, delivered deadpan, becomes a proposal. Sarcasm becomes agreement. Thinking aloud becomes a plan. Somebody says "we could just cancel the whole project" to make a rhetorical point, everyone laughs, and there it sits under DECISIONS with a perfectly straight face.
The things it cannot see at all
Notetakers listen. Mostly, they do not watch. So a meeting whose real content was on a whiteboard, in a shared screen, in a chart everyone was pointing at, or in a diagram someone sketched, arrives as a transcript full of the word "this".
"So if we move this here, that fixes it" is a perfect record of a sentence and a useless record of a meeting. If the value of your session lived in something visual, photograph the whiteboard, keep the slides, and treat the summary as a supporting note rather than the record.
The same goes for the room itself. Nods, folded arms, the person who visibly did not agree but said nothing, the long pause before someone answered. None of that reaches the transcript, and quite a lot of it was the meeting.
Checkpoint
A notetaker captures audio, converts it to text, then summarises that text, so an error early in the chain gets repeated later with total confidence.
When the argument was the point
Some meetings exist to produce a conclusion, and a summary serves those very well. Other meetings exist so that four people can disagree properly, and the value is in the reasoning: the objection someone raised, the thing that was tried two years ago and failed, the reason the obvious option is not the right one.
Summaries flatten all of that. You get a neat outcome and lose the shape of the debate, which is precisely the part you will want in six months when somebody new asks why on earth you chose this. For those meetings, a summary is not enough. Ask specifically for the objections raised and the options rejected, and be prepared to write a paragraph yourself.
| Meeting type | Notetaker fit | Why |
|---|---|---|
| Weekly status update | Good | Repetitive, factual, low stakes, people mostly speak in turn |
| Project kick off | Good | Lots of detail nobody will remember, plenty of actions to extract |
| Client call | Good, with care | Useful record, but consent matters more and errors cost more |
| Whiteboard or design session | Poor | The content is visual and the transcript fills up with "this" and "that" |
| Heated strategy debate | Poor | The reasoning matters more than the conclusion, and summaries flatten reasoning |
| One to one about performance | Do not record | Sensitive, personal, and recording changes what gets said |
Some meetings should not be recorded at all
This is the honest bit, and it belongs in lesson one rather than buried at the back of the course.
Do not put a notetaker in a one to one about someone's performance. Do not record a grievance, a disciplinary conversation, a redundancy discussion, a conversation about somebody's health or their family, or any meeting whose subject is a named person who is not in the room. Do not record the meeting where a colleague is likely to cry.
There are two reasons and the second is the larger one. The first is that a permanent, searchable, easily forwarded record of a sensitive conversation is a liability, and it is your liability. The second is that recording changes the meeting. People do not speak honestly into a microphone they can see. The moment a bot joins a difficult conversation, you stop getting the truth and start getting the version that is safe to have written down. You will have gained a tidy summary of a conversation that was no longer worth having.
If you catch yourself thinking "I probably should not record this, but the notes would be handy", the hesitation is the answer. Leave the notetaker out and write four lines by hand afterwards.
Checkpoint
Visual meetings, meetings whose value was the argument, and anything sensitive about a named person are the wrong jobs for a notetaker.
๐ Quiz
Question 1 of 3A summary calls your colleague 'Ravi Sharman' when he is actually Ravi Sharma. Which stage caused it, and why does the summary repeat it so confidently?