Can an AI Notetaker Record an In-Person Meeting?

The short answer

Yes. Almost every tool on a normal shortlist now has some path to it, and the paths are different enough that the answer to can it is less useful than the answer to what gets worse. Transcribing four people around a table is a solved problem in the sense that you will get readable text back. The parts that degrade are the parts you do not notice until you go looking for the note two months later.

There are two ways a notetaker can hear a meeting, and only one of them was ever able to walk into a room. Knowing which one you are buying explains most of what follows.

A bot cannot enter a room

A bot-based notetaker works by accepting a calendar invite and joining a meeting link as a participant. That is the whole mechanism. Take away the link and there is nothing to join, so the entire product stops at the door of a conference room.

The vendors built around that model solved it the only way available, by shipping something that listens to a device microphone instead. Fathom's own overview page puts in-person capture on its phone app: capture in-person meetings and review summaries, notes, and insights on the go, all from the Fathom iOS app. Otter's help centre gives two options, recording through a browser on a computer or recording on a phone or tablet through its app, and says that by default Otter will pull from your device's built-in microphone and that the audio quality and what it captures will depend on your device's audio input.

The other family never had the problem. A tool that reads audio from your own machine is already in the room, because your laptop is in the room. Granola's documentation describes exactly this, no meeting bot, running only on your computer and using your system audio and microphone, and it also ships phone apps. Driffle works the same way on the Mac and treats a call and a table as the same event.

So the capability question is largely settled. The buying question is what a single microphone in a room does to the output, and that is where the two families stop looking alike.

What degrades first is who said what

On a video call, every participant arrives with a name attached by the platform. The notetaker does not have to work out who is talking from the sound of the voice, because Zoom already knows. In a room, none of that exists. Four people share one microphone and one mixed track, and the only thing separating them is acoustics.

Granola's documentation is unusually direct about the consequence. Speaker tags, it says, are available for Google Meet, Zoom, and Microsoft Teams on macOS and Windows. Where speaker tags are not available, you get Me and Them, which is your microphone on one side and everything else on the other. Their phone app can recognise different speakers face to face, but the laptop app in a conference room has no participant list to read, so a four person conversation flattens into two labels: you, and the room.

That sounds cosmetic and it is not. A commitment without an owner is not an action item, it is a sentence. If the transcript says the launch is moving to November and cannot say whether that came from the person who controls the launch date or from someone speculating, you have a record that reads well and cannot be acted on. Six weeks later, when two people remember it differently, the note settles nothing.

The cheapest fix costs nothing and no vendor will tell you about it, because it makes their software look weaker. Say names out loud. Open with who is in the room. When you hand over, use the person's name. Every AI summariser in this category infers speakers from context when it cannot hear them apart, and names in the text are the context. It feels stilted for about a minute.

Two ordinary settings that quietly ruin the audio

Both of these are in vendor documentation and neither appears in any roundup of tools for in-person meetings.

Neither of these shows up as an error. You get a note either way. It is just a note about a shorter, quieter meeting than the one you were in.

• Your headphones. Otter's own FAQ says it plainly: bluetooth headphones can be great for recording your own voice, but not other speakers around you. If your earbuds are connected, your laptop's audio input is a microphone sitting in your ear, pointed at your jaw. The room gets quieter, not louder. Take them out for a conversation happening in front of you, and put the laptop between people rather than in front of you. Otter's guidance is to place the microphone in the centre of all speakers, or near the speaker you want to capture.
• Your laptop lid. Granola's documentation lists the situations in which it stops transcribing, and one of them is when you put your computer to sleep. Any tool that listens through a Mac has the same property, because a sleeping machine has no microphone. In-person is exactly the meeting where closing the laptop feels like the polite thing to do. Leave it open, on power, and expect that the record ends wherever you shut the lid, which in practice is often right before the part worth keeping.

The half nobody sells you: the meeting has no invite

Here is the failure that costs the most and gets the least attention. A calendar event does an enormous amount of unpaid filing work. It carries a title, a start time, an attendee list, a recurrence pattern, and usually enough in the invite text to say which account or project the conversation belongs to. Every notetaker on the market leans on that metadata to organise what it captures, because it is free and it is already correct.

A conversation at a table supplies none of it. You press record, you talk for fifty minutes, and what lands is a transcript with no title, no participants and no relationship to anything else in the archive. Granola's docs describe one small symptom of this, that ad-hoc notes started from a new note have no scheduled end time and so the tool cannot reliably tell when the meeting ended. The larger version of the same problem is that the note has no handle. In three weeks you will remember the conversation clearly and be unable to name the day it happened, and the search box wants a word from a transcript you never read.

Then the part that should change how you weigh all of this. The conversations that happen in rooms are not a random sample of your meetings. They are the customer visit, the dinner where an investor says the actual objection, the offsite where the roadmap genuinely changed, the ten minutes in the corridor after the board meeting. Remote teams hold their routine business on video and travel for the conversations that carry weight. So the category of meeting your system indexes worst is, on average, the category carrying the most decision weight per minute.

Which means the right test of a notetaker for in-person work has almost nothing to do with the microphone. It is whether a conversation with no calendar event attached can still be found later by someone who only remembers what it was about.

How to test it in a trial, at a table

Desk tests pass everything. One person, a quiet room, a laptop eighteen inches away, and every tool in this market produces a clean transcript. The failures only appear at a table, so run the trial at one.

The first five take twenty minutes. The last one is the one that decides it, and it is the only one you cannot rush.

• Use four people and a real table, with one person at the far end who speaks twice and quietly. Read the transcript for those two moments specifically.
• Talk over someone deliberately for ten seconds, then read that minute. Overlapping speech on one channel is where single-microphone capture actually breaks, and it is normal in a room.
• Run one meeting with your headphones in and one without, and compare what the tool heard of everyone else.
• Close the lid for two minutes in the middle. See whether the record resumes, ends silently, or splits.
• Check what the note is called and who it says was present, before you edit anything. That is what the archive will hold.
• Wait three weeks. Then find that conversation by asking a question about what was decided, without scrolling a list and without remembering the date. Most tools fail here, and this is the failure that matters.

Where Driffle sits, including the limits

Driffle transcribes audio from your own device, so a Zoom call and a conversation across a table go through the same path and land in the same place. Nothing joins the call, and in a room there is nothing to join anyway. Because every capture lands in one searchable memory alongside the documents and screens you were working in, the room conversation is not filed in a separate app with different rules, and you can ask a question later and get the answer with the moment it came from attached.

The limits, stated plainly, because a trial will find them anyway. Driffle is a Mac app for Apple Silicon running macOS 14.4 or later, downloaded directly rather than from the App Store. There is no phone app today, so for in-person the Mac has to be in the room and awake. Transcription and the AI run in the cloud, so it needs a connection and there is no on-device or offline mode. The audio itself is transcribed in real time and discarded rather than stored, which is deliberate, and it means there is no recording to replay afterwards.

If most of your important conversations happen in rooms and your current notetaker has never seen one of them, Driffle is a free download for Mac at driffle.ai/download.