AI Agent
Guardrails: checks before the model
Blocked terms, allowed topics, and what the assistant must never see.
AI AgentRead it in the appThe same article opens from Help inside app.dotask.io, beside the screen it describes.
- Blocked terms — words or phrases the assistant must not engage with. A message containing one gets the reply you write here, and the model is never called.
- Allowed topics — describe what the assistant may discuss; a short classifier checks each message first and off-topic ones get your off-topic reply.
- Personal data — what the assistant may know about the customer (name, email, which fields), and card-number masking so a pasted card number never reaches the model.
Guardrails run before anything is retrieved or spent, so they cost nothing and cannot be talked around. Rules under Behavior → Rules sit under them, never above.