Without a boundary, AI improvises exactly where it should not
With no explicit line, a model faced with a complaint, a medical question or an out-of-policy discount will still produce something that sounds reasonable. Sounding reasonable is the risk: it will not tell you it was unsure.
Judge on risk, intent and complexity
Hand over immediately on risk (health, refunds, legal). Hand over on high intent with unusual conditions (large orders, custom work, urgency). Hand over after three rounds with no progress. The three are independent; any one of them is enough.
Hand over the context, not just the customer
A handoff is not passing someone to a colleague who asks everything again. Send the confirmed need, what has already been provided, and where it is stuck. By the third time a customer repeats themselves, the odds of closing have halved.
Making a customer repeat themselves three times costs more than a wrong answer.
Record why each handoff happened
The distribution of handoff reasons is an honest map of where the AI's competence ends. A reason that keeps recurring is knowledge or a rule to add next round — not headcount.