Ask people what they dislike about company chatbots and the answer is rarely that the bot was wrong. It is that after five minutes of explaining, they reached a human and had to explain it again.
The handover is where most of the goodwill is lost, and it is usually the least designed part of the system.
Pass the transcript, and pass a summary
The minimum is that the agent can see the full conversation before saying anything. Surprisingly often they cannot, because the bot and the helpdesk are separate systems that were integrated only to the extent of opening a ticket.
The transcript alone is not enough though. An agent picking up a live conversation has seconds, not minutes, and a forty-turn log is not readable in seconds. What works is a short structured summary at the top: what the customer wants, what has been established, what has already been tried, and what the bot was unable to resolve.
That summary is a genuinely good use of a model, and it is the single highest-value thing to add to most existing deployments.
Pass what the bot already collected
If the customer gave an order number, an account reference or a description of the fault, those should arrive with the conversation and be visible in the agent’s interface without a search.
Asking again is the specific moment customers conclude the bot was a waste of their time. It also signals that the two systems are not really connected, which undermines confidence in everything else you have automated.
Decide when to stop trying
Containment targets create pressure to keep the bot engaged for one more turn. That is exactly backwards once a conversation has gone wrong.
Sensible triggers to hand over without further attempts:
- The customer asks for a human. Once, clearly, and it should happen — not after a retention loop.
- The same intent fails twice. A third attempt almost never works.
- Detected frustration, or language suggesting a complaint rather than a question.
- Anything touching money, cancellation, a complaint with legal implications, or a safety issue.
- The conversation exceeds a turn limit without progress.
A bot that hands over quickly and cleanly is experienced as helpful. One that resists is experienced as an obstacle, regardless of how good its answers are.
Be honest about what happens next
If a human is available now, say so and connect. If the handover creates a ticket that will be answered tomorrow, say that instead — and say it before the customer waits.
The worst pattern is an ambiguous “connecting you to an agent” that resolves into silence. People will wait a long time if they know they are waiting; they will not forgive being left uncertain.
Out of hours, state it plainly with an expected response time rather than letting the conversation trail off.
Make the return path work
Handover is usually treated as one-directional, but the agent often resolves something the bot could have handled next time — or needs to hand back for a routine follow-up.
Two things worth building: a way for the agent to flag a conversation as something the bot should have handled, which becomes your improvement backlog; and continuity so that if the customer returns tomorrow, the history is there rather than starting fresh.
What to measure
Not containment. Handover quality is measured by whether the customer had to repeat information, how long they waited between the handover and a human response, and how often agents report arriving without usable context.
Ask the agents directly. They know exactly which handovers are useless and they are rarely asked.
Common questions
What should be passed to a human agent during chatbot handover?
The full transcript, plus a short structured summary of what the customer wants, what has been established and what was already tried — and every identifier the bot collected, visible without a search.
When should a chatbot hand over to a human?
Immediately on a clear request, after the same intent fails twice, on detected frustration, on anything involving money, cancellation, legal complaints or safety, and when a conversation exceeds a turn limit without progress.
Why do customers dislike chatbots?
Usually not because the answers are wrong, but because reaching a human means explaining everything again. Most of the frustration is created at the handover rather than during the conversation.
Should containment rate be the target?
No. Containment targets create pressure to keep customers engaged with the bot one turn longer, which is exactly wrong once a conversation has gone badly. Measure whether customers had to repeat themselves and how long they waited for a human.
Our Call Center Agent escalates with context by design. See AI chatbots and virtual agents, or read how to measure containment rate honestly.
