The wrong goal for a support chatbot is full automation. The right goal is narrower: answer the repeatable questions instantly, and get everything else to a human while the buyer still cares. Most "our chatbot failed" stories are not model failures — they are handoff design failures. The bot answered something it shouldn't have, looped on a question it didn't know, or dumped the buyer into a queue nobody watched.
Four triggers that must hand off
Explicit request for a human. "Let me talk to a real person" is not an objection to overcome — it's an instruction. The bot should confirm and hand off immediately. Arguing with this request burns trust faster than any wrong answer.
Money and liability. Refunds, disputes, custom orders, anything where the answer commits the shop to spending money or changes an order. The bot can explain the policy, but the decision goes to a human. A chatbot that improvises refund terms is a liability generator.
No confident match. When nothing in the knowledge base matches with reasonable confidence, say so and hand off. A confident wrong answer is worse than an honest "let me get someone who knows this one" — the buyer forgives a handoff, not being misled.
Negative emotion. Repeated complaints, frustration phrasing, refund threats. At that point the conversation has stopped being a question and become a situation. Route it.
What the bot does before handing off
- Collect context once. Order number, question type, what the buyer already tried. One short round of questions, not an interrogation.
- Never loop. If the buyer rephrases the same question twice without getting a usable answer, stop answering — a third attempt at the same entry reads as stalling.
- Say what happens next. Who replies, in which channel, roughly when. A silent handoff — "please wait" into a queue that never moves — reads as being ignored.
What the human agent should see
The handoff arrives pre-contextualized or it hasn't really happened: full transcript, detected intent, confidence score, and the knowledge-base entries that were tried. The buyer should never have to repeat their question, and the agent shouldn't have to scroll forty messages to reconstruct it. This is the part that decides whether "AI first, human second" actually saves time — the bot's instant first reply only pays off if the pickup is fast too.
Failure modes worth designing against
The retry loop. The bot rephrases the same wrong answer three times, then offers the human option. By then the buyer has typed the question three ways and is done. Hand off on the second miss, not the fifth.
The fake human. A bot with a human name that denies being a bot. It always gets discovered, and the discovery retroactively poisons every answer it gave. Say what it is.
The invisible queue. "Connecting you to an agent…" with nobody on the other end and no stated hours. If the human team is offline, the honest version is: working hours, plus a leave-a-message path with a real SLA.
Handoff as punishment. Treating the human transfer as a failure state to be avoided at all costs, so the bot fights for it for ten minutes first. The buyer is gone by then. The four triggers above are early exits, on purpose.
What to measure
- Handoff rate — what share of conversations needed a human. A healthy number is not zero; zero means the triggers are too tight or the bot is dodging.
- Self-service resolution for the repeatable set, tracked separately from the handed-off set.
- First response time, where the bot's instant reply counts. That instant response is the entire point of bot-first.
- Re-open rate after handoff — did the pickup actually resolve it, or did the buyer come back angrier.
If the only KPI is automation rate, the system learns to avoid handoffs, and you get confident wrong answers at scale. Measure resolution and handoff quality as a pair.
How we run it
We build customer-support auto-reply tooling for e-commerce sellers under Dingdang CS, and the bot-to-human handoff rules above are written up in detail in our handoff configuration guide — including the trigger list, the pre-handoff context capture, and how the queue side is configured so pickups stay fast.
Top comments (1)
the four triggers read right. one more i'd add is time: if the bot hasn't resolved it in a set number of turns, or the human hasn't picked it up within a set number of minutes, escalate again to a second named person. the failure you describe, a queue nobody watched, needs a timer as well as a trigger.