Managing Approvals
Handle approval requests triggered by lethal-trifecta detection or policy rules.
When an AI action requires oversight, Edison Watch pauses the tool call and sends an approval prompt. The prompt arrives on whichever channels you've enabled in Settings → Approval notifications - answering on any one of them dismisses it everywhere else.
Two things can trigger an approval prompt:
- Lethal Trifecta detection - the session has accumulated all three risk flags (private data access + untrusted content exposure + external communication).
- A policy rule with
require_approval- an admin-defined CEL rule matched the call and requested human sign-off. See Policy Rules.
Delivery Channels
| Channel | How it arrives |
|---|---|
| Web app | Popup in the Edison Watch dashboard (enabled by default). |
| Desktop app | Native notification from the Edison Watch desktop app (enabled by default). |
| Slack | Message with Approve / Deny buttons, sent to a configured channel or DM. |
| Telegram | Message with Approve / Deny buttons, sent to a linked Telegram chat. |
Configure and test your channels in Settings → Approval notifications. Each enabled channel has a Send test button so you can verify delivery before relying on it.
Approving or denying on any channel immediately dismisses the prompt on all other channels. The decision is recorded in the session's audit trail.
Making a Decision
✅ Approve
The AI is doing what you asked. The action proceeds.
❌ Deny
The action is suspicious or unrequested. The AI is blocked from completing this specific task.
The AI is paused: Your assistant cannot continue until you respond. If nobody answers within the configured timeout, the call is denied by default (fail-closed). A rule can opt into fail-open with approval_on_timeout: allow.
Reviewing Context
If you're unsure, check the session's history in the Dashboard:
- Go to Sessions.
- Find the active session.
- Review the timeline to see what data the AI has accessed.
Tips for Safe Approvals
- Check the destination: Is it sending data where you expect?
- Verify the tool: Is the AI using the correct tool for your request?
- Trust your gut: If an action seems unnecessary, deny it.