Field notesCustomer & commerce
Handing a customer from bot to person without making them repeat it.
The handover decides how customers remember the automation. Escalate on clear triggers, pass a summary with the data already gathered, and tell the customer honestly what happens next.
veridive5 min read
The customer has explained the problem to the bot twice. The bot finally hands over, and the agent’s first message is “Merhaba, size nasıl yardımcı olabilirim?” The customer explains it a third time, to a person who seems never to have heard of it. Whatever the automation did well before that moment, this is what the customer will remember.
The handover is where automation earns or loses trust. Escalate on clear triggers, give the person a card with everything already gathered and said, and tell the customer honestly what happens next.
Why do handovers go wrong?
Three failures come up again and again, and none of them is about staffing:
- The context stays behind. The agent receives a ticket number, or a raw transcript to scroll through, and asks the customer to start again.
- The bot lets go too late. It offers the same help article after the customer has said twice that it doesn’t help.
- The customer hears something untrue. “Connecting you now”, followed by a long silence, or a promise the agent then has to take back.
When should a conversation be handed over?
Write the triggers down as rules, each with an owner:
- A sensitive request: complaints, payment problems, legal threats, anything about health or safety.
- Low confidence: no source supports an answer, or the request is still unclear after one clarifying question.
- Negative sentiment or an explicit request: the customer asks for a person in any wording (“temsilci”, “yetkili”, “a real person”) or is clearly frustrated.
- A repeated question: the same thing asked again usually means the answer didn’t help.
- High value: business accounts, large orders or amounts above a threshold the owner sets.
Done looks like: every handover logs which trigger fired, so the rules can be tuned. What goes wrong: relying on a sentiment score alone. “Harika, üçüncü kez yazıyorum” (Great, I’m writing for the third time) reads as positive to a naive model, so combine signals: the words, the repetition and the history of the case.
What should the person receive?
A handover card on one screen, with the transcript one click away. It needs six fields: the request type, the customer and order data, what the bot told the customer word for word, what it already checked, why it handed over, and a suggested next step. “What the bot said” matters more than it looks: without it, the agent can contradict an answer the customer already has.
In an illustrative case, a customer writes for the second time about a refund for a returned jacket. The agent receives this card:
| Field | In the illustrative case |
|---|---|
| Request type | Refund not received, second contact |
| Customer and order | Identity verified; return received at the warehouse; refund issued to the original card |
| What the bot said | “Your refund has been issued to your card. Your bank decides when it appears.” |
| What was checked | Return status and refund status in the payment system; no failed transactions |
| Why handed over | Repeated question, and the customer asked for a person |
| Suggested next step | Share the refund reference for the customer’s bank; if the bank sees nothing, open a payment investigation |
The agent’s first message can now continue the conversation: the refund went out, here is the reference your bank will need, and here is what happens if they can’t find it.
Done looks like: agents can act on the card without scrolling the transcript. What goes wrong: a card built from whatever is easy to export, instead of from the questions agents otherwise have to ask again.
The handover is the part of the automation customers remember.
What should the customer be told?
Three things: that a person is taking over, roughly when they will reply, and that they won’t need to repeat themselves. “Konuyu bir çalışanımıza aktardım. Yazdıklarınızı görecek, tekrar anlatmanıza gerek yok.” (I’ve passed this to a colleague. They will see what you wrote, so you won’t need to explain again.) Base the wait on the real queue, not a hopeful constant, and once the handover starts, keep the bot quiet in that conversation so it doesn’t answer over the agent. What goes wrong: a fixed promise such as “within five minutes” that the queue can’t keep on a busy day.
What if no one is available?
Out of hours, or when the queue is full, say so plainly. Tell the customer when the team is back, and offer a callback or a ticket number they can quote. Keep the card, so the first agent on shift starts with the full context rather than an empty ticket. Decide in advance which cases can’t wait, such as a payment taken twice, and where they go. What never works is a dead end: a loop back into the bot, or “please try again later” with nothing attached.
How do you measure handover quality?
Measure the handover itself, separately from the automation:
- Repeated questions after handover: how often the customer restates the problem or the order number after the agent’s first message. Check a sample of transcripts every week.
- Time to first human reply after the handover, by channel and hour.
- Customer effort: one question at the end, such as how easy it was to get the issue solved.
- Trigger mix: which triggers fire, and which handovers agents mark as unnecessary.
A rising share of low-confidence handovers often points to a gap in the knowledge base, not to the model.
Design the card from real handovers
Read a few dozen handovers from the agent’s side of the screen. Note what the agent had to ask again, and turn those questions into the fields of your card. The handover is where customer operations design meets system design, and building the card into your tools is typical custom AI software work. If you are still deciding what to automate at all, start with chatbot or agent assist.
Ask an assistant about this note