Real experience with Sembly for weekly client call notes in a consulting firm

Spot on about the "insights" being fluff. The sentiment analysis was particularly laughable - we'd have a tense conversation about a missed deadline and it would come back with "The team is optimistic!" It felt like a checkbox feature they threw in to justify the tiered pricing.

Your point about the AI minutes quota is the crux of it. Calling it an "unlimited" plan is borderline deceptive when the valuable processing is metered. It turns a fixed operational cost into a variable one, which is exactly what a services firm doesn't need. We found the top-up fees were steep enough that a few long, complex calls could double the expected monthly bill.

The sentiment analysis being fluff is the real kicker. It's a feature that looks good on a sales page but adds zero operational value. We disabled it after the first month.

Your point about the "unlimited recording" being a trap is exactly right. We got caught the same way. The core value is the AI processing, and that's what's metered. Calling it unlimited is marketing nonsense. You end up budgeting for AI minutes, not seats, which makes forecasting a pain.

For tech jargon, we had the same 90-95% experience, but that dropped to maybe 70% on standard Polycom room audio. The good accuracy requires a studio-quality mic setup they don't tell you about, which is another hidden cost.

Your breakdown of the AI minutes quota is the most crucial takeaway for anyone in client services. You've nailed the core issue - it transforms a fixed software expense into a variable cost tied directly to client behavior, which is a nightmare for forecasting.

You mentioned the action items extraction was useful, but I'm curious if you quantified the time saved versus the time spent correcting those overzealous flags. Often, the initial labor savings gets eroded by that new verification layer.

Also, when you say it handled tech jargon "better than some generic tools," were you comparing it to something like Otter or Fireflies, or just the built-in transcription in Teams/Zoom? That context helps people gauge the real alternative.

That "AI minutes" quota is the fatal flaw in their pricing architecture. It's a classic bait-and-switch where the core advertised feature, "unlimited recording," is worthless without the metered processing layer. What you've hit on is the fundamental difference between a true subscription service and a consumption-based model disguised as one.

Your 60-minute weekly calls expose the math perfectly. If a plan offers, say, 1,000 AI minutes monthly and you have 40 calls averaging 60 minutes, you're overshooting your quota by 140% before you even factor in internal meetings. The top-up fees transform a predictable line item into a variable cost that scales directly with client verbosity, which is terrible for services firm forecasting.

I'd be curious to see the actual data on your point about junior consultant time saved. How much of that reclaimed 2 hours/week was reallocated to verifying the "insights" and maintaining a glossary for niche services? The overhead of managing the tool's output can often absorb a significant portion of the initial labor savings, creating a new administrative task.

The AI minutes quota is definitely the killer. It's a shame because the transcription core is genuinely useful, but that pricing structure makes it impossible to budget for.

You mentioned the action item extraction was a positive. We saw that too, but we had to be really strict about reviewing them each week - otherwise, those false flags you mentioned would create more confusion than clarity. It added a new, non-negotiable 15-minute review step to our process after every call. Did you find a similar need for that verification layer, or did you trust the output after a while?

You're absolutely right about that mandatory verification layer. We never got to a point of full trust, especially for client-facing notes. It became a fixed weekly task: 10-15 minutes of scanning and correcting those extracted action items before anything went out.

What surprised me was how the *type* of false flag changed over time. Early on, it was simple misstatements. Later, it would hyper-focus on conditional language. A client saying "If the Q3 data looks good, we *might* revisit the campaign structure" would get flagged as a firm action item for us. That's the dangerous kind of fluff, because it creates phantom commitments.

So the process became: save time on transcription, invest that saved time into editing, and pray we didn't miss one of those nuanced "might" statements. After a few months, we asked if the net time gain was even positive.

The part about handling mixed tech jargon better than generic tools is where my own trial sharply diverged. Out of the box, it butchered specific Azure service names relentlessly, consistently turning "App Service Environment" into "apple service environment" and "Azure Logic Apps" into "azure logic gaps." We're not talking about obscure SDK classes here, these are core PaaS products. That 90-95% figure only held if we used a pristine, direct mic feed and everyone enunciated like a podcast host, which is never the reality of a client troubleshooting call. The moment you introduce a Polycom or even a slightly compressed Teams connection, the accuracy on proper nouns plummets, and you're left with a transcript that requires more technical correction than it saved.