Supervisor guardrail
Supervisor guardrail
A second check on every reply about to be sent automatically; anything that breaks your policies waits for an agent.
With the supervisor guardrail on, every answer that would be sent automatically — e-mail and help-desk replies, chat-widget answers, SMS / WhatsApp / Messenger messages — gets a second check first. If it fails, the reply is not sent: it is held for review like any reply below the auto-send threshold, and the reason is shown in the decision trace.
| Check | How |
|---|---|
| Personal data | Payment card numbers, bank accounts (IBAN), social security numbers, secret keys and other people’s e-mail addresses — checked instantly, no AI call |
| Your brand voice | “Words to avoid” from Brand voice — checked instantly |
| Competitors | Your competitor names from AI settings — checked instantly |
| Promises and advice | Refunds, credits, compensation, guarantees, legal / medical / financial advice, admitting fault — one small AI call |
| Your escalation topics | Topics your escalation rules send to people (legal threats, cancellations …) — same AI call |
- 1Turn it onDashboard → Trust & controls → Supervisor guardrail.
- 2Watch the review queueHeld replies appear in the review queue with the reason. Approve, edit or reject them as usual.
Note:Cost: about one small AI call per automatically sent reply. If the check itself fails (for example the AI service is overloaded) the reply is held, never sent unchecked.
Tips & tricks
- Fill in “Words to avoid” in Brand voice and your competitor names first — those checks are free and instant.
- Turn it on for the first weeks of a rollout or after a big knowledge-base change, when you want extra safety.
- If the same kind of reply keeps being held, fix the article it came from or add an escalation rule so it goes straight to a person.