The summary is the version that exists

The bot joined before I did. I clicked into a strategy call last month with a group I advise informally, and it was already there in the participant list, "Otter Notetaker," recording, before anyone had said a word. Nobody remarked on it. Forty minutes later, meeting over, a summary arrived in the shared channel: five bullet points, three action items, a two-line overview written in a neutral, capable voice. The bullet points were accurate as far as I could remember the meeting, which is the problem. I can't fully remember the meeting. None of us can. The summary has quietly become the version of the meeting that exists, and it arrived already organized, which is a property I've learned to distrust in documents.

The scene is the argument. A meeting used to produce a memory that belonged to the people in it, and now it produces a record, generated by a third party, that belongs to whoever holds the account. Everything about Otter.ai, good and bad, falls out of that shift, and the interesting part is not the accuracy. It's which document people end up believing.

The summary is a claim wearing a record's clothes

What the tool does is transcribe meetings in real time, then compress them: key points, action items, an overview, all extracted without being asked for. In hands-on testing by Jovana Simić at The Business Dive in April 2026, transcript accuracy ran around 85 percent, below the competitors tested in the same sessions, and the errors had a specific character: the transcription tool turned "Fireflies," the name of a rival product, into "Fireplace," and "best lead generation tools" into "best regeneration tools," and once identified a third speaker in a two-person call (the review is here). Those are small failures in a transcript, because a transcript invites checking. The same failures in a summary become decisions, because a summary invites circulation. The reviewer watched it extract "book flights by Friday" from a mock meeting without prompting, which is the feature working exactly as designed, and it's also the risk in miniature: extraction is interpretation, an action item is a claim about what a person committed to, and the person who never said it will have to notice a bullet point in order to object. Nobody opens the transcript to check a bullet point. The transcript is right there, and that's what makes the arrangement so stable.

The bot attends for the account that pays

Consider the incentive, because the business structure explains the consent structure. Otter charges per seat, on meters of minutes: the free plan covers 300 minutes a month with a 30-minute cap per recording, Pro is $16.99 monthly or $8.33 billed annually for 1,200 minutes, Business is $30 monthly or $19.99 annually for 6,000 minutes and the admin controls, and no tier on the ladder is unlimited.

A company paid per minute of meeting processed has every reason to make capture automatic, and automatic is what it is: the bot joins scheduled calls off your Google Calendar, announces itself, and records. The consent announcement goes to everyone in the room, which is more than many tools bother with, and the consent itself belongs to the host. Attendees on Reddit complain about exactly this, being recorded by someone else's subscription, and the complaint is well-founded even when the tool follows its own rules: your meeting is the raw material of somebody else's archive, and the value of that archive accrues to the account, not to the people speaking in it. The language support tells you who the customer is. English, Spanish, French, nothing else, in a year when transcription in thirty languages is a commodity. That's not a technical limit. It's a market decision, and the market is the English-language enterprise meeting.

The strongest case: an index, not a verdict

The counterargument deserves its best form, and here it is. Eighty-five percent accuracy is a catastrophe for a record and plenty for an index. The transcript exists to make the audio searchable: you jump to a timestamp, press play, and hear the person say it themselves, in their own voice, and the audio never degrades. For the colleague who was double-booked, the hard-of-hearing colleague the tool genuinely serves, the second-language speaker who wants to recheck a fast exchange, the journalist finding the exact quote in two hours of tape, the searchable archive of everything a team said is a real asset, and the alternative is a human note-taker whose reliability is unauditable and whose memory degrades from the first minute. The failure I described at the top is not the tool's doing. A bullet point becomes a decision only where a team has decided, lazily, to treat summaries as the record, and the discipline is available to any team that wants it: attach the transcript, keep the audio, treat the summary as a cover page. The category has also heard the objection. A crop of bot-less note-takers, Granola and Krisp and Radiant among them, now transcribes locally on your own machine without joining the call as a participant, which is the market pricing the consent problem rather than dismissing it.

The variable is what circulates

It depends, and the dependency is whether the summary functions as memory or as the record. For personal recall, finding what was said, Otter at $8.33 a month is a fair index with a real archive behind it. For decisions with consequences, hiring, budgets, commitments made in a meeting someone will act on, the rule I'd hold is that a summary never circulates without its transcript attached, and the audio stays until the action item is dead. Watch one variable over the next couple of years: whether summaries start forwarding themselves by default, out of the tool and into the channels where work is judged, because that's the moment the 15 percent error rate becomes everyone's memory. The bot was recording when I arrived, and the summary was believed before I objected. I used to think the fix was better transcription. The fix is knowing which document the room is actually reading.