AI handover
When AI agents escalate to a human, and how to make the handoff invisible to the customer.
Handover is when an AI agent stops and gives the conversation to a person. Getting it right matters more than getting the AI's answers right — an agent that answers 90% well and escalates the rest cleanly beats one that attempts everything.
When agents hand over
| Trigger | Why |
|---|---|
| Low confidence | The agent isn't sure enough to answer. |
| Customer asks for a person | Always honour this, immediately. |
| Sensitive topic | Payment details, account changes, refunds, complaints. |
| Repeated failure | The agent has tried and the customer is still stuck. |
| Configured limit | A turn or time limit you set. |
Configure these per agent in AI Agents.
Make "talk to a human" work
If a customer types "agent", "human", or "speak to someone", the agent should hand over on that message — no clarifying question, no one more attempt.
Customers who ask for a person and get another AI reply escalate to being annoyed customers. The cost of an unnecessary handover is a few minutes of an agent's time; the cost of trapping someone is a bad review.
You can reinforce this with a Keyword Received trigger on AGENT or HUMAN that assigns the conversation directly.
What a human receives
A handed-over conversation arrives in the inbox with the full AI exchange visible. Your agent can read what was already asked and answered, and continue without making the customer repeat themselves.
Repeating the customer back to themselves is the most common way a handover goes wrong, and it's entirely avoidable — the context is right there.
The handover trigger
Handover fires the AI Handed to Human workflow trigger, with {{trigger.handover.reason}} available.
This is what makes handover operational rather than just a state change:
Trigger: AI Handed to Human
→ Condition: {{trigger.handover.reason}} contains "refund"
→ Add / Remove Tag: "urgent"
→ Assign to Agent: senior supportOther useful patterns: notify a team when handover happens outside business hours, tag by reason so you can see what your AI can't handle, or route billing questions to a specific person.
Handover and billing
A handed-over conversation is not an AI resolution, and you aren't charged for it.
You're only charged when the AI solved the problem and closed the conversation itself. Handover, human takeover, and 24-hour idle auto-close are all free.
That means over-escalating costs you nothing directly. Tune towards handing over too readily rather than too rarely.
Reviewing handovers
The reasons agents hand over tell you what to fix. Look for patterns:
| Pattern | What it means |
|---|---|
| Same question repeatedly | Add it to the knowledge base. |
| Handover right after greeting | Instructions are too cautious, or knowledge is too thin. |
| Never hands over | Check it's escalating on sensitive topics — silence here is a risk, not a win. |
| Handover on a topic you documented | The source may not be Ready, or its wording doesn't match how customers ask. |
After handover
A human-owned conversation stays human-owned. The AI doesn't resume mid-thread.
If the customer starts a new conversation later, the AI handles it as normal.
Next steps
- AI agents — configuring agents.
- Knowledge base — reducing unnecessary handovers.
- Inbox — where handovers land.