Automatic transcription is the most successful thing generative AI has produced for ordinary office workers, and almost nobody treats it as an achievement. When OpenAI released Whisper as open source in September 2022, accurate speech-to-text stopped being a product and became a commodity. Zoom, Microsoft Teams and Google Meet all now offer some form of AI notes, and standalone tools like Otter and Fireflies had been doing it for years before that. It works. That’s what makes the next part uncomfortable.

The summary is the wrong deliverable

Every one of these tools leads with the summary: five bullets and a list of action items, delivered before you’ve closed the call. It’s tidy, and it’s also a lossy compression of a conversation by a system that doesn’t know which throwaway comment the project actually hinged on. A human note-taker picks up the hesitation in someone’s voice when they say “that should be fine.” A summarizer writes down “Timeline confirmed.”

The transcript is the valuable artifact. It’s searchable, it settles arguments about who said what, and it can be re-queried months later with questions nobody thought to ask in the meeting. But transcripts are long and ugly, so products bury them behind the neat recap, and teams read the recap and nothing else.

What attendance turns into

Once a tool is promising to catch everything, people stop catching anything. I’d bet most readers have been on a call where half the participants were visibly elsewhere, trusting the bot. That’s an observation, not a measured result, but it matches how every automation of attention has gone, from spellcheck to GPS.

The cost doesn’t show up when the summary is right. It shows up on the one call in twenty where it’s subtly wrong, and nobody present remembers well enough to object.

A bot that joins a call and records everyone is also a legal question. Recording consent rules vary by jurisdiction, and in some US states all parties must agree. Most tools handle this with a banner or a bot name in the participant list, which is a weak form of consent. Companies deploying these at scale should have a written policy, and most of the ones I can think of don’t.

A better default

If I were designing these products, the transcript would be the primary view, with timestamps and speaker labels, and the summary would be a collapsible layer on top. Highlights would link back to the exact moment in the recording, so checking a claim takes ten seconds instead of a forty-minute replay. A few tools already link summary lines to the transcript, which is the right instinct. It should be the baseline, not a premium feature.