You already know: the support agent (actions under policy), the web agent (untrusted content trying to steer an agent) and the voice assistant (privacy boundaries). Email combines them: a huge volume of private data, an attacker can send content straight into the system simply by emailing the user, and sending is irreversible.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud; prices are illustrative.
1. The prompt, and the questions that change the design
A mid-level answer: "summarise each email and generate a reply". A staff candidate sees four harder problems: volume (most email does not need a reply, so drafting everything wastes money), voice (drafts must sound like the user without training on everyone's private mail), inbound injection (anyone can email instructions to the assistant), and irreversibility (a sent email cannot be recalled).
Assumptions you state: