You already know: the voice assistant (streaming speech recognition), long-form synthesis (claims tied to sources) and the email assistant (drafts the user approves). A meeting copilot combines them, with two problems of its own: who said what (an action item assigned to the wrong person is worse than none) and consent (recording people is regulated, and not everyone in the call is your user).
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative. Recording and consent rules vary by jurisdiction; this lesson describes design patterns, not legal advice.
1. The prompt, and the questions that change the design
A mid-level answer: "transcribe, then ask a model to summarise". A staff candidate sees four harder problems: speaker attribution (who agreed to do what), consent and access (who was told, who may see the summary), cost (transcription of hundreds of thousands of hours a day), and actions (follow-ups and tasks must be approved, not fired automatically).
Assumptions you state: