A capable AI receptionist still needs a clear answer to one question: when should it stop being the receptionist and bring in a person?
Hand off when authority runs out
Escalate decisions the system is not authorized to make: unusual pricing, refunds, guarantees, disputes, safety judgments, legal or medical conclusions, and exceptions to policy.
Hand off when context is fragile
Strong emotion, repeated misunderstanding, accessibility needs, background noise, language limitations, or conflicting account details can make continued automation worse than a concise transfer or message.
Respect the caller's request
If a caller clearly asks for a person, the flow should acknowledge it and follow the approved path. That path may be a live transfer during office hours and a priority callback message after hours; it should not be a loop disguised as help.
Make the transfer survivable
Pass the caller's name, reason, urgency, and collected details when the system supports it. If no one answers, return to a safe fallback and preserve the context. Test the failed-transfer experience as carefully as the successful one.
Review thresholds with evidence
Sample real handoffs and missed handoffs. Tighten rules that interrupt people without value, and widen rules where callers are getting stuck. Human oversight is a design choice, not an admission that the system failed.