How do I stop the agent saying certain things?

Open Workspace → AI Agent → Knowledge → Never say. The Never say list is a set of hard limits. Anything on it is refused in every conversation, on every channel, whatever the customer asks and however they ask it.

The list ships with reviewed defaults. They are limits that were decided on, not suggestions to consider — leave them unless you have a specific reason.

How a guardrail differs from a fact

They pull in opposite directions and it is worth keeping them straight:

• A fact is something the agent may actively tell a customer — your turnaround, your payment terms. See What can the agent promise on my behalf?.
• A guardrail is something the agent must never say, whether or not it could work the answer out.

A guardrail wins. If a fact and a guardrail could both apply, the agent refuses rather than answers — the limit is bound harder in its instructions than the knowledge is.

Writing a good guardrail

Write the sentence you never want sent, not the topic you are nervous about.

• Works: "never promise a delivery date on a Sunday"
• Too vague: "be careful with timing"

A vague guardrail gives the agent nothing to refuse against. A specific one is unambiguous both to the agent and to a colleague reading the list later.

What happens when a customer asks anyway

The agent declines that part of the answer and carries on with the rest of the conversation. It does not go silent, and it does not invent a reason. If the question genuinely needs a person, it hands the conversation over — see What does the AI agent do, and what does it know?.

Removing a default

You can remove any entry with Remove. Removing a guardrail does not make the agent start saying that thing — it only stops it being refused outright, so the agent falls back to your facts and documents. If nothing there covers it, it still will not invent an answer.