We rolled out an AI note taker across our org about six months ago and the reaction has been mixed in ways I didn’t fully anticipate. The initial pitch was straightforward: better meeting documentation, less time spent writing up action items, more focus during the call itself. All of that is basically true for standard conversation in quiet environments.
The accuracy problems show up in specific conditions that turn out to be pretty common in practice.
Technical terminology is the obvious one. Our product team runs meetings with vocabulary that the transcription model clearly wasn’t trained on. Names, product features, industry-specific terms. The model approximates these with phonetically similar words that are wrong in ways that matter. A transcription that says ‘customer churn’ when someone said a specific product name is more confusing than no transcription at all.
Accents are the bigger issue and the one I feel more careful about raising. We have a genuinely international team. Several colleagues whose first language isn’t English produce transcriptions with noticeably higher error rates. When that means their contributions are less accurately captured in the meeting record, that’s not just a technical problem. It’s a fairness one.
The consent piece also came up. We’d assumed everyone was fine with it once we announced the rollout. We hadn’t asked individually. A few people raised concerns after the fact that we should have surfaced earlier.
We still use the tool, but with more explicit guidance about when to turn it off and more expectation-setting about accuracy limitations. Curious whether other people-ops folks have navigated the language and accuracy fairness question specifically.