That 20% maintenance buffer isn't a luxury, it's a necessity. People always forget the quiet, costly failures: the API changes its webhook signature, your parsing regex breaks because someone said "K eight S" instead of "Kubernetes," and suddenly your CRM is filling up with nonsense for two weeks before anyone notices.
The vendor lock-in point is key, but it cuts both ways. Building your own abstraction layer over the transcription API adds even more maintenance. I've seen teams spend more time keeping their "transcription service adapter" up to date than they ever saved by avoiding a bundled tool.
Your last line is the gospel truth. If the output doesn't land where the work already happens, it's digital shelfware. We pipe summaries into the Jira ticket or PR the discussion was about, not into a separate "insights" app. Otherwise, it's just another forgotten tab.