You join a video call and a small tile appears next to the participants. "Notetaker." Nobody invited it, exactly — someone's calendar assistant did, weeks ago, and it has been quietly showing up ever since. Halfway through, a colleague mentions a client by name, then a salary figure, then a mistake someone made last quarter. The bot is still sitting there. Nobody says anything, because saying something would be awkward and everyone assumes somebody has already checked.
Almost nobody has checked. AI note takers went from novelty to default in about two years, and the adoption curve badly outran the part where organizations decide what the rules are. The tools themselves are genuinely useful — that's why they spread — but the useful part and the risky part come from exactly the same mechanism, and it's worth understanding what that mechanism actually is.
Transcription stopped being the hard part. What happens to the recording afterward is the hard part.
What's actually happening when the bot joins
Under the hood, most of these tools do the same four things in sequence, and the differences between products live in where each step happens.
First, capture. The tool needs the audio. There are two broad approaches, and the split matters more than any feature list. A bot-based tool joins your meeting as a participant — it appears in the roster, it's visible to everyone, and it receives the meeting's audio and often video stream the same way a human attendee does. A botless or on-device tool instead captures system audio locally on your machine, so nothing extra appears in the participant list.
Second, transcription. Speech becomes text, usually with speaker labels attached. This step is now so commoditized that accuracy is table stakes; the major tools are all broadly competent on clear English audio, and all of them degrade on crosstalk, heavy accents, and bad microphones in roughly similar ways.
Third, summarization. A language model turns the transcript into structured output: a summary, decisions, action items, sometimes a follow-up email draft. This is where the products genuinely differ in quality, because it depends on prompting, on how well the model handles long context, and on whether the tool knows anything about your team's vocabulary.
Fourth, storage and distribution. The transcript, the summary, and often the raw recording get saved somewhere and pushed somewhere else — a shared workspace, a CRM record, a Slack channel, an email.
Steps one through three are the product. Step four is the risk, and it's the step people evaluate least.
The three questions that actually matter
Feature comparisons are mostly noise. If you're choosing a tool — or auditing one your team already uses — these three questions separate them.
Where does the audio go? In the common cloud model, your meeting audio is uploaded to the vendor's servers, processed there, and stored there. In the on-device model, capture and sometimes processing happen locally, and only the resulting text may be synced. Neither is inherently wrong; a cloud tool with strong controls can be perfectly appropriate for routine internal standups. But you cannot make a sensible decision about client calls, medical conversations, HR discussions, or anything under a confidentiality agreement without knowing which model you're using. The answer is in the vendor's documentation, and it is worth ten minutes of reading before it becomes a question someone asks you under pressure.
Is your content used to train models? Vendor policies here vary widely and — this is the important part — they change. Some tools state plainly that customer meeting content is not used to train their models. Others reserve the right to train on de-identified transcripts or audio. Some offer it as a toggle that may be on by default. Because these terms get revised, treating any specific claim as permanent is a mistake; check the current policy at the time you deploy the tool, and check it again when you renew.
Who can see the output, and for how long? A transcript that lands in a channel visible to the whole company is a very different object from one visible only to attendees. Look for per-meeting sharing controls, retention settings you can actually configure, and a real deletion path. "We keep everything forever by default" is a defensible product choice and an indefensible surprise.
The consent problem nobody wants to own
In many jurisdictions, recording a conversation without the participants' knowledge is not merely rude. Rules differ substantially — some places require only one party to consent, others require everyone, and cross-border calls can implicate more than one regime at once. This is genuinely a legal question rather than a technical one, and if your meetings routinely cross borders or touch regulated information, it deserves an actual answer from someone qualified rather than a guess from an IT ticket.
The practical layer underneath the legal one is simpler, and you can fix it this week:
- Announce it in the first ten seconds. "I've got a notetaker running, shout if you'd rather I didn't." It costs nothing and it removes the low hum of unease that otherwise sits under the whole call.
- Have an off switch you'll actually use. The moment a conversation turns to a person's performance, health, compensation, or a legal dispute, the note taker should leave. If turning it off is annoying, it won't happen.
- Decide who gets the transcript before the meeting, not after. Especially with external participants.
- Treat botless tools as requiring more disclosure, not less. The convenience of an invisible capture tool is exactly what makes announcing it more important. A visible bot at least discloses itself.
One under-discussed side effect: people talk differently when they know they're being recorded. The half-formed idea, the honest "I'm not sure this is working," the candid read on a stalled project — these are the things that stop being said. If a recurring meeting exists specifically to surface problems early, an always-on note taker may quietly make it worse at its job. That's not a privacy issue. It's a much more expensive one.
Where they're worth it, and where they aren't
They earn their keep in recurring meetings with clear outputs — sprint planning, project syncs, sales discovery calls, customer support escalations, interviews where you'd otherwise be typing instead of listening. The single biggest gain is not the summary. It's that the person who would have been taking notes gets to participate instead. In a five-person meeting, that's a 20% increase in the number of people actually in the room.
They're also quietly excellent for accessibility — searchable transcripts help people who process text better than speech, who joined a call in a second language, or who missed the meeting entirely and would otherwise have booked a 30-minute "catch me up" call with someone busy.
They're a poor fit for conversations whose value is in the room rather than the record: one-on-ones about how someone's doing, difficult feedback, early brainstorming that needs people to say dumb things safely, anything involving personnel decisions or legal exposure. In those, the tool doesn't just add risk — it changes the conversation.
There's also a failure mode worth naming: transcript sprawl. A team can end up generating thousands of pages nobody reads, which is harmless until the day it isn't, because that archive is now discoverable, searchable by anyone with access, and a large surface area in the event of a breach. Retention settings exist for a reason. Six months is a defensible default for most routine internal meetings; forever rarely is.
A checklist you can run in twenty minutes
If a note taker is already in your meetings and nobody has ever formally evaluated it, work through this:
- List every tool in use. These spread through individual sign-ups; there is often more than one, and IT may know about none of them.
- For each, find the data location and the training policy in the current documentation, not in a blog post from last year.
- Check the default sharing scope. Open the last transcript and ask who could see it.
- Set a retention period and confirm deletion actually deletes rather than hides.
- Write two sentences of policy — which meetings are allowed to be recorded, and which are not — and put them where people schedule meetings, not in a wiki nobody opens.
- Add the announcement to your own meeting habit and let it spread by imitation.
None of this requires a procurement process or a committee. It requires one person spending twenty minutes on questions that currently have no owner.
The short version
AI note takers are one of the rare productivity tools that deliver roughly what they promise. The transcription works, the summaries are usable, and the meeting genuinely improves when nobody is heads-down typing. The technology is not the problem.
The problem is that the tool is a recording device that arrived without any of the social norms recording devices normally come with. Nobody would set a voice recorder on a conference table without mentioning it, and yet the software equivalent joins hundreds of thousands of calls a day with no announcement at all.
Know where the audio goes. Say it out loud when it's running. Turn it off when the conversation gets human. Do those three things and you get most of the benefit and very little of the exposure — which is a better trade than most technology offers.
Vendor capabilities and data policies described here are general patterns as of writing and change frequently; verify current terms with any tool before deploying it. Recording laws vary by jurisdiction — this is general information, not legal advice.
