Hey all. New here, and still trying out different tools at my company.
We do a weekly all-hands that's recorded, and I've been asked to look into Sembly for generating meeting notes and summaries. The demos look good, but I'm curious about real, weekly production use.
Anyone using it consistently for this? Specifically wondering:
- How accurate are the action items over time?
- Does it handle recurring meetings with mostly the same attendees well?
- Any major pitfalls when the recording quality isn't perfect?
We're a SaaS/B2B team, so the context is usually product and project updates. Just trying to see if it's reliable enough to build into our workflow.
We've been using Sembly for our engineering standup recordings for about four months now, so I can speak to recurring meetings. The action item accuracy is... decent, but not perfect. It's great at catching explicit "I will do X by Friday" statements, but sometimes misses implied tasks when someone says "I'll look into that." We've gotten into the habit of phrasing things more clearly for the AI, which isn't a bad thing actually.
>Any major pitfalls when the recording quality isn't perfect?
This is the biggest gotcha. If someone's on a poor connection or there's background noise, the transcript gets patchy and the summary can go off the rails. We make sure the host records locally as a backup and uploads that file, which helps a ton.
For product updates context, it works well once it learns recurring names and project codenames. Give it a few weeks to build up that meeting history.
Prompt engineering is the new debugging
We've been running it for quarterly planning offsites, not weekly all-hands, but the accuracy principles translate. For >How accurate are the action items over time?, our data shows a 70-80% capture rate on explicit action items, but that's only when the audio source is pristine. The real issue is that it often assigns the action to the wrong person if multiple voices are similar, or if someone speaks over the assignee saying "I'll take that."
The biggest pitfall for a weekly all-hands is the summary coherence. When you have 30+ people and a wide-ranging agenda, the generated summary often tries to force a narrative thread that wasn't there, conflating different product updates. You'll need a human to heavily edit the output; it's not a fire-and-forget tool.
For your use case, I'd recommend a dual-source approach: use Sembly for the initial transcript and to flag potential action items, but have a designated note-taker validate and structure the final summary. It saves time but isn't autonomous.