The wrong way to measure a support agent is deflection rate. Optimize for "tickets the human never saw" and you'll build something that stonewalls frustrated customers to protect its own metric. The right measure is resolution — including the conversations where the right resolution is a human.
Here's how we think about the handoff.
Escalate on signal, not just on ask
Waiting for a customer to type "let me talk to a person" is too late — by then they're already annoyed. The agent watches for earlier signals:
- The question is outside the configured scope or tools.
- Retrieval keeps coming back empty for what they're asking.
- Sentiment turns — repeated rephrasing, frustration, urgency.
- The action carries real risk and policy says a human signs off.
Any of these can trigger a handoff before the customer has to demand one.
Hand off with context, not a cold start
A handoff that dumps the customer into a fresh queue to re-explain everything is barely better than no agent at all. When Bund AI escalates, your team receives the full transcript, what the agent already tried, the customer record, and why it escalated — so the first human reply moves the conversation forward instead of restarting it.
Write actions always confirm
Reading information is safe; changing it is not. Look-ups run inline, but anything that mutates a real system — booking, refunding, ticketing — pauses for explicit confirmation. The agent proposes; a person (or the customer) approves. Nothing irreversible happens on a model's say-so alone.
The best automation knows the shape of its own competence — and steps aside cleanly at the edge.
Make the seam invisible
To the customer, the goal is one continuous conversation. They shouldn't feel the moment the agent steps back and a teammate steps in. Shared context is what makes that seam disappear — and it's where an AI agent earns its place in your support stack.