1. Documentation
  2. AI support
  3. Configuring

Escalation and guardrails

When the agent hands over, and who it does not answer.


Escalation

Escalation is what separates a useful assistant from a polite wall.

  • Escalate when the answer is unknown (on by default): rather than invent, the agent calls a human.
  • Roles to notify: who gets mentioned.
  • Escalation message: what the member sees.
  • Attach a summary: the staff arrives with the problem and what has already been tried, instead of reading it all back.

Bao does not merely ask the model politely to escalate. Every reply is reviewed automatically: if it leaves the member without a solution, Bao escalates on its own, whatever the model decided. An empty reply is treated as the worst case, not the mildest.

Guardrails

GuardrailEffect
Allowed rolesEmpty = everyone. Otherwise, only these roles can call on the agent
Blocked rolesThe agent never answers them
Forbidden topicsThe agent declines and points to a human
Refusal messageWhat it replies in that case

Allowed roles are used in particular to reserve an expensive agent for your paying customers.

👍 / 👎 votes

On by default, under each reply. They are not decoration: they feed the detection of knowledge gaps.