Feature · Trust & safety

Honest about what I know. And what I don't.

The guardrails are structural — built into how I answer, not features you can disable. Citation, vision honesty, self- correction, persona consistency, no deflection.

Five guardrails

What the customer can count on.

• Cite every answer

  Every reply is traceable to a source in your knowledge base. The customer hears the answer; your dashboard shows you the doc section it came from. When I don't have a source, I say so cleanly instead of fabricating.

• Vision honesty

  I always know whether I can actually see the screen versus inferring from docs. When labels are unclear, I describe position instead of guessing. I never invent a button that isn't on the page.

• Self-correction

  If a user pushes back on something I said and I was wrong, I admit it. I don't double down on a hallucination. The transcript captures the correction so your reviewers can see what happened.

• Stays in character

  I never say 'as an AI' or break the persona mid-conversation. The voice and tone you configured is the voice and tone the customer hears. End to end.

• Never deflects

  I answer from my own knowledge instead of sending users to docs. If I know it, I tell them. If I don't, I say so and route the question — but the default isn't 'check the docs'.