How approval cards work
When the AI proposes a destructive or surprising action, it sends an inline card you can Approve or Cancel — nothing happens until you tap.
When you ask the AI for something destructive or surprising — “delete these tasks”, “mark all leave requests approved”, “notify Animesh on the bell” — it doesn’t just go do it. It posts an approval card in the chat thread with the tool name, the affected items, and two buttons: Approve and Cancel.
Tap Approve and the action runs server-side. The card flips to a green check with a one-line summary (“Approved leave request”, “Marked 3 tasks as done”). Tap Cancel and the card greys out to “Skipped” — the AI moves on without acting.
Why two tiers
Cheap, reversible actions (list, get, search) and your-own- scope writes (create a task for yourself, set a personal reminder, log a contact note) run automatically without a card. Anything that touches other people's data, deletes something, or affects HR records gets the approval gate.
Who can approve?
Only the user who initiated the chat turn. The cards are locked to your account — even if a teammate happens to see your screen, they can’t approve on your behalf.