“Can I talk to a human?” is the most important sentence on an AI call. Handled well, it turns suspicion into trust. Handled badly — or ignored — it turns a caller into a complaint.
We treat escalation as part of the protocol, not a fallback. It can be triggered by intent, keyword, or policy, and it hands the human agent full context so the caller doesn't have to start over. Every handoff is logged.
The result is counterintuitive: making it trivial to leave the AI is what makes people willing to stay.