The real split in meeting AI: a bot in the call or audio on your device

Most meeting-tool comparisons I read are feature spreadsheets: who integrates with which CRM, who has the prettier action-item view, who bills by the seat. Useful once you've already chosen. Useless for the choice that actually shapes everything downstream, and that choice is architecture, not a feature. Does a bot join your call as a participant, or does software on your own device listen to the audio the call is already producing? Nearly every trade-off that follows is downstream of that one fork.

Pick that axis wrong and no feature list saves you.

the two camps, plainly

On one side, the bot-based tools. Otter, Fireflies, Fathom, tl;dv and its kin, Read.ai, Avoma, Sembly, Fellow. The shared mechanism is a participant that joins the meeting, records the stream, and hands it to the cloud for processing. Because that bot lives server-side, it wires naturally into calendars and CRMs, and it keeps recording whether or not your laptop happens to be awake.

On the other side, the on-device tools. Jamie, Granola, Krisp. These capture the system audio locally, put no participant in the attendee list, and at least begin their processing on your machine. I wrote up the on-device experience through Jamie specifically; Granola plays a related card as an augmented notepad, Krisp arrived from the noise-cancellation side. Different products, one architectural stance.

what follows from that one choice

Consent optics come first, because people react to them in the room. A visible bot in the attendee list announces that recording's happening, which is honest but occasionally chills the conversation. Silent on-device capture is discreet, which feels comfortable right up until you notice that discretion isn't the same thing as permission. Where the legal weight actually sits is its own article, where your meeting audio goes; here I only want to flag that the two architectures hand you opposite social defaults for free.

In-person coverage is the cleanest divider. A bot can't join a table in a cafe, because there's no meeting link to join. On-device capture doesn't care whether the sound came from Zoom or from the person across the table. Half my sensitive conversations happen face to face, so this one property alone rules out an entire camp for part of my week.

Then there's where the audio flows. Bot-based means the stream is cloud-bound by design; that's the product, not a leak. On-device means the raw audio at least starts on your hardware, and depending on the tool may never leave it, or may sync only the text. And there's the trade you pay for custody: the cloud tools, precisely because they live server-side, tend to offer deeper integrations than the on-device camp manages. Custody and convenience pull against each other here, and honest vendors don't pretend otherwise.

The bot-versus-device split is really one question wearing a marketing costume: whose machine holds the audio first, yours or theirs.

the failure modes are mirror images

Each camp fails in the exact shape its architecture predicts, which is oddly reassuring once you've been burned by both.

A bot fails by not joining. Wrong meeting link, stuck in a waiting room, a host who never admits it, a locked meeting. I once sat through a forty-minute call whose bot idled in a waiting room the entire time, and I walked away with precisely zero notes for the one conversation that week I actually needed on the record. When a bot misses, it misses the whole thing.

On-device fails by going quiet. The laptop sleeps, a battery dies, the capture app loses the audio device mid-call, and you get a summary with a hole in it that nobody labels as a hole. That's the sneakier failure, because a bot that no-shows is obvious, whereas a five-minute gap in an otherwise clean transcript looks exactly like a stretch where nothing was said.

how I actually pick

My rule after a year of this is unglamorous. Big recurring internal calls, where the org already consented once and I want the notes flowing into our tools, lean bot. Client 1:1s, anything sensitive, anything face to face, lean on-device. And when a compliance clause rules out both commercial camps at once, there's a third door, which is to roll your own from Whisper and a local model.

Whichever camp writes the first draft, a human still has to read it before it becomes a decision of record, which is the ordinary human-in-the-loop discipline these tools quietly depend on and rarely put on the box. The architecture decides where your audio lives and how your meetings feel. It doesn't decide whether the notes are right. You still do.