AI human handoff playbook
When AI should answer, when it should ask, when it should page a person — and how to design the agent payload so replies are fast and accurate.
AI overconfidence is the #1 reason customers escalate
Most chatbot complaints come from confident wrong answers. Customers can tolerate AI not knowing — they cannot tolerate AI inventing.
The fix: design the routing layer so AI is conservative by default. When in doubt, ask. When stakes are high, hand off. When the customer says "human", hand off without asking.
Configure these on day one
- Customer asks for a person. Direct request, frustrated language. Hand off immediately.
- Confidence below threshold. Sonar score under your configured value. Offer a person.
- Policy keyword. Refund, complaint, legal, medical, billing. Force handoff regardless of confidence.
- Customer tag. VIP, high-LTV, named accounts. Bypass AI by policy.
- Tool failure. Knowledge lookup or live tool errored. Hand off — never fake an answer.
- Buying intent. High-cart-value or qualified-prospect signals route to sales.
A handoff that respects the agent's time
The agent should be able to reply within 30 seconds of being paged. To get there, the handoff payload needs:
Two-line summary. What the customer wants, what AI tried, why escalation fired.
Customer profile. Name, channel, language, prior history, lifetime value.
Suggested reply. Templated response in your tone, ready to send or edit.
Transcript link. Full conversation, AI reasoning, tool calls.
Lock state. Confirm the agent owns the conversation. No double replies.
What to avoid
- Confidence-only routing. Pair with explicit triggers.
- Silent handoff. Tell the customer a person is joining; don't just disappear AI.
- No release path. Agents need to release a conversation back to AI cleanly.
- Fake availability. If no one is on call, say so. Propose a callback time.
- Hidden audit. Every handoff goes into the audit trail with reason and actor.
Frequently asked questions
What is the most common handoff mistake?
Trusting AI confidence as the only signal. Confidence captures uncertainty about the answer, not customer expectation. Pair confidence with explicit triggers (keyword groups, policy categories, customer attributes).
How do I tune the confidence threshold?
Start conservative — high threshold means more handoffs. Sample escalations weekly for the first month, drop the threshold once you trust the routing. Always keep policy triggers above the confidence layer.
How do I avoid losing conversations during handoff?
Use a per-tenant routing layer with explicit locks. Only one endpoint owns a conversation at a time. The platform handles this in the tenant state layer.
Keep exploring
Design your handoff with us.
Bring your current routing rules. We'll review them and propose specific changes that reduce escalations and improve agent response time.