Human-in-the-loop approval is a design pattern in which an AI system prepares an action but waits for a person to review and approve it before the action takes effect. In email, the typical checkpoint is the send button: the AI drafts the reply, and a human decides whether it goes out as written, edited, or not at all.
Where the approval checkpoint sits
The useful question is which actions need a human and which don't. Reading a message, searching files, or checking a calendar changes nothing anyone else sees, so most systems let the AI do those alone. Sending a message, accepting an invite, or sharing a file speaks for you, so that's where the checkpoint goes.
A good approval step shows the context next to the proposed action (the original message, the draft, the attached file) so the review takes seconds instead of a reread of the whole thread.
Human-in-the-loop vs. human-on-the-loop
In a human-in-the-loop setup, the action doesn't happen until someone approves it. In a human-on-the-loop setup, the system acts on its own and a person monitors and can step in afterward. An email tool that auto-sends replies and lets you review the sent folder is on-the-loop, even if it calls the feature "review."
How Synced uses human-in-the-loop approval
Synced drafts replies but never sends one without approval. Every AI-drafted message waits in the Decide queue until you approve or edit it. The reasoning behind that choice is in Why Synced asks before it sends.