The handover from bot to human is where most chatbots fail
Automation quality is judged at the moment it stops. A clean escalation makes a limited bot feel competent; a bad one makes a capable bot feel like an obstruction.
Key takeaways
- Escalate immediately on explicit request, low confidence, repeated fallback, or emotional and high-stakes content.
- The agent must inherit identity, transcript and any retrieved records; asking the customer to repeat themselves turns the interaction hostile.
- State a specific reply time when no agent is available rather than making a vague promise.
- Measure satisfaction split by whether the conversation escalated; it tells you whether to expand automation or fix the handover.
Customers do not judge a chatbot on how much it automates. They judge it on what happens when it cannot help. A bot that resolves forty per cent of conversations and hands the rest to a person cleanly, with full context and no repetition, is experienced as good service. A bot that resolves seventy per cent and traps the remainder in a loop is experienced as a company avoiding its customers.
Four triggers that should always escalate
- Explicit request. If someone asks for a human, escalate immediately and without negotiation. Making a customer ask three times is the single most damaging pattern in conversational support.
- Low confidence. When the system is not reasonably sure it understood, say so and offer a person rather than guessing. A confident wrong answer costs more than an admission.
- Repetition. Two consecutive fallbacks, or the same question rephrased twice, means the conversation is not converging. Escalate on the second, not the fourth.
- Emotional or high-stakes content. Complaints, cancellations, payment disputes and anything expressing frustration. These need judgement, and the cost of getting one wrong exceeds the saving from automating a hundred routine ones.
What must travel with the customer
The agent should open the conversation already knowing who the customer is, what they asked, what the bot attempted, and any record already retrieved such as an order or an account. Asking a customer to repeat what they just typed is the moment the interaction turns hostile, and it is entirely avoidable. The full transcript should be visible in the agent's interface, not summarised into a ticket subject line.
Honesty about availability
If no agent is available, say so with a specific expectation rather than a vague promise. Outside business hours, state when someone will reply and then meet it. In the region this matters more than teams assume, because messaging feels synchronous to the customer; a message that goes unanswered for fourteen hours without explanation reads as being ignored, whereas the same delay with a stated return time reads as normal.
Never end a conversation by asking the customer to call a number or send an email. They chose this channel. Redirecting them to another one is a refusal dressed as help.
Measuring the handover specifically
Track escalation rate, time to first human reply after escalation, and satisfaction split by whether the conversation was escalated. That third measure is the useful one: if escalated conversations score well, your automation boundary is correctly placed and you can push it outward. If they score badly, the handover itself is broken and expanding automation will make things worse.
Customers judge a bot on what happens when it cannot help, not on how much it automates.