chandima
← writing / 04

when an agent should hand over

2025-11-28 · ai · 3 min

The reliability ceiling of an AI customer agent is rarely the model — it is knowing the exact turn to step aside for a human.

Demos reward agents that keep talking no matter what the user asks. Production punishes them. One confident wrong answer about a refund policy undoes fifty fast automated replies.

01 / make escalation a first-class tool

Instead of treating handoff as a fallback error state, model it as a first-class action with explicit triggers: policy exceptions, billing disputes, low retrieval grounding, or direct requests for a person.

When the agent hands over, it writes a compact three-line state summary for the human operator — what the customer wants, what was already verified, and which boundary triggered escalation — so the customer never repeats themselves.

“A great agent solves eighty percent quietly and hands over the last twenty percent without ego.”