The automated part is almost never the problem. The customer asks, the agent answers, the conversation moves. The friction shows up a second later: the agent decides this one needs a human, opens a ticket titled “customer question”, and the person on the other side opens with “what is your name?”.
A good hand-off carries three things along with the conversation: the full transcript, whatever has already been verified (identity, order, plan) and the reason for the transfer. If any of the three fails to cross, the customer repeats everything, and repeating is the single largest source of dissatisfaction in an automated channel, larger even than waiting.
The three triggers worth having
- An explicit request. “I want to talk to someone” is not negotiable and should not be met with a round of persuasion.
- Repeated failure. Two answers in a row where the agent found nothing to ground a reply is signal enough. Waiting for the third only sours the mood of whoever picks the conversation up.
- Detected frustration. It works better as queue priority than as a trigger on its own, because angry text is not always aimed at the agent.
What almost never works as a trigger is conversation length. A long conversation can be detailed research going well. Length is not a symptom.
What has to cross with it
- The full transcript, not just the summary. The summary is what the human reads fast; the full text is what prevents the second identical question.
- The data already confirmed: email, order number, plan, whatever the conversation validated. Asking again says the first ten minutes did not count.
- The reason for the transfer, in one line. “Asked for a refund outside policy” lets the human open the conversation already knowing the argument.
If the human has to open with “tell me what happened”, the hand-off did not happen. The conversation simply restarted somewhere else.
The topics that should never be automated
Refunds, formal complaints, cancellation with retention, and anything with legal weight. Not because the model would get the answer wrong, but because the cost of being wrong is asymmetric: getting it right saves two minutes, getting it wrong becomes a case. Route those topics straight to a human, before any attempt at an answer.
Once
The hand-off that comes back
The other direction gets discussed less and matters just as much. Once the human has resolved it, the conversation should return to the agent instead of sitting in an open queue. Follow-up questions (delivery date, status, a duplicate invoice) are exactly the kind of thing it answers well, and nobody needs to stay on standby for them.
And outside business hours
A hand-off assumes someone is on the other side. At three in the morning nobody is. The honest behaviour is to say so: record the request, state when someone will answer, and make sure the conversation arrives there with everything already gathered. Promising a transfer and delivering silence is worse than saying the team is back at nine.
In the end the hand-off is the seam between what automation resolves and what it should not attempt. The definition lives in the glossary, and if the decision is still agent versus human team, the comparison with a human SDR sets out where each one wins.
