Exception Inbox
Approve or reject the AI agent actions that require a human decision, with priority, deadlines, and a full-context detail view.
Exception Inbox (/settings/autonomy/escalations) collects the "Actions requiring human
approval" — the moments an AI agent hits a hard override, a money cap, or a case it may not act on
alone (see Autonomy Settings). The subtitle counts how many are pending, and a clear
inbox shows "No pending escalations. All actions are within approved boundaries."
Triage an item
Each escalation carries a priority badge — P1 — Immediate, P2 — High, P3 — Medium, or P4 — Low — plus its action category, a Deadline, the Default action if you do nothing, and an Overdue badge once the deadline passes. Act directly from the row:
- Approve — let the agent carry out its recommendation.
- Reject — block it.
For the full picture, open the detail dialog (Escalation Details, "Review the full context before making a decision"), which shows the Recommendation, the Evidence, and any Alternatives. From here you have three choices: Approve, Modify (approve a changed version), or Reject.
See resolved items
By default the inbox lists pending work. Toggle Show Resolved (which then reads Showing All) to include escalations you have already handled. Long lists paginate with Previous and Next.