Half of reviews end in "yes, but change one thing." Edit with AI keeps that inside the queue: open the action, describe the change the way you would tell a colleague, and the draft rewrites in place.
What works well: tone ("warmer", "more formal", "less apologetic"), structure ("shorten to three sentences", "lead with the deadline"), and substance ("use the pallet price from Friday's email", "add a response deadline of Thursday", "adjust the total to $4,920"). Each edit produces a new draft you read before approving; you always approve the exact text that sends.
What to watch: substantive edits to amounts and terms update the connected records consistently (the invoice and its cover email move together), and the action's risk level re-evaluates if an edit changes the amount. If an edit request is ambiguous, the draft asks rather than guesses.
Edits are also telemetry: your edit rate falling over the first weeks is the main signal the system is learning your voice and context. If a specific correction keeps recurring, mention it in onboarding; that is usually a missing piece of context we can connect.