Design a human-in-the-loop approval gate that doesn't create rubber-stamping

Task: agent drafts and sends outbound partner emails. Approval UI shows: the recipient, the specific claim or commitment in the email (e.g. 'promises a refund of $X' or 'commits to a delivery date'), and a diff-highlighted view against the standard template, not the full draft prose, so the reviewer's eye goes straight to what's non-standard. Uncertainty surfacing: if the agent had to infer the recipient's preferred name or title from ambiguous context, flag that specific line with a yellow marker, since that's where errors actually occur, not the boilerplate closing paragraph. Approval fatigue fix: since you expect ~200/day and most are template-standard, auto-send emails that pass a strict template-match check with zero flagged deviations, and route only the ones with a flagged deviation (expected to be a small minority) to human review, this keeps the gate meaningful for the cases that actually need it instead of 200 daily rubber-stamps. Rejection: human provides a one-line reason, agent uses it to redraft once and re-submit, doesn't auto-retry silently. Audit log records: original draft, flagged deviations, reviewer decision, reason if rejected, and timestamp, kept separately from the agent's own action log so review quality can be audited independently later.