Boundaries, sensitive topics, and human handoff

Spell this in custom instructions and mirror key facts in the brief: no made-up pricing, timelines, SLAs, customer names, integrations, certifications, or policy exceptions. When data is missing, the right move is to say so and point to a canonical page or a human.

Regulated claims (health outcomes, financial advice, legal interpretation) should default to “we cannot advise; here is how to reach our team or your professional.” Err on the side of deferral.

For topics with high emotional or legal stakes, instruct the agent to acknowledge, gather only what is needed, and route to a human channel you monitor. Provide the exact path: email, form URL, or support portal—not vague “contact us.”

• Billing and refunds — quote only what is in the brief; offer human review for edge cases.
• Security — never ask for secrets, passwords, or full card numbers in chat; explain how official verification works.
• Abuse or harassment — a short, firm policy plus escalation beats debate in the widget.

Good handoffs summarize what you already know (“You are on Pro and need SAML—our team can confirm rollout timing”) and set expectations (“We reply within one business day”). Bad handoffs dump visitors on a generic form with no context.

If sales owns demos, say so. If support owns bugs, separate the path. Confusion about ownership is worse than no chat at all.

Do not put confidential roadmaps, unreleased features, or internal metrics in the knowledge brief to sound impressive. If it is not public and approved for customers, it should not be in the agent's grounding.