Chatbot to human handoff, designed so nobody repeats themselves.

A good handoff has three parts. The bot knows when to stop. The person who takes over knows what already happened. And the customer knows what’s going on the whole time. Miss any one of them and the handoff fails, even if the routing works perfectly.

Think of the handoff as a feature of the bot, not an admission that it failed. Every bot meets questions it shouldn’t answer, customers who want a person, and moments where a wrong answer costs real money. The design question isn’t whether to hand off. It’s how early, how cleanly, and to whom.

The test is simple. Read a handed-off conversation from the customer’s side. If they had to repeat their name, their order number or their problem, the handoff needs work.

Most handoffs come from five kinds of trigger. Set all five, not just the obvious one, because customers rarely type “I want a human” before they get frustrated.

Explicit request. If the customer asks for a person, a manager or an agent, hand off. Don’t make them ask twice, and don’t try one more answer first.

Low confidence. Your bot platform scores how sure it is about each answer. Set a confidence threshold with your vendor and tune it on real transcripts. Too low and the bot guesses; too high and it hands off questions it could have answered.

Repeated failure. When the bot misses on the same intent more than once, stop looping. After two failed attempts is a sensible starting rule; tune it once you see real conversations. Frustration is the fifth trigger: all-caps messages, swearing, “this is useless”, or the customer repeating the same question word for word.

Scope and business rules. Some topics should go to a person no matter how confident the bot is: payments, refunds and disputes, cancellations, legal threats, medical or safety questions, complaints about staff, and anything involving a vulnerable customer. Write this list down with whoever owns risk in your business.

Hiding the human option is the most common way bots damage trust. Customers who can’t find an exit start typing nonsense to break the bot, and some simply leave. Put a “Talk to a person” button or menu option in every conversation, not only after the bot gives up.

This doesn’t mean everyone will take it. Most people try the bot first when it’s quick and useful. The visible exit is what makes them comfortable trying. On voice, the same rule applies: saying “agent” or “representative” or pressing zero should always work.

The agent needs two things the moment the conversation arrives: the full transcript and a short summary on top of it. The transcript is the record. The summary is what the agent actually reads in the few seconds before they reply.

Keep the summary short and structured, so every agent knows where to look. Generate it automatically, but check a sample of summaries against transcripts every week, because a confident wrong summary is worse than none.

Silence after “Connecting you to an agent” is where handoffs fall apart. Tell the customer three things: that a person is taking over, roughly how long it will take, and that they don’t need to repeat themselves.

Be honest about the wait. Show a real estimate from your queue if your platform has one. If it doesn’t, say “a few minutes” only if that’s usually true, and offer a way to leave a message instead of waiting. A short wait you promised and kept beats a long one you never mentioned.

When the person arrives, have them introduce themselves by name. The customer should be able to tell the bot has left and a human has joined.

A handoff with nobody on the other end is worse than no handoff at all. Before launch, decide exactly what the bot does when no one is staffed.

The usual pattern: the bot says plainly that the team is offline, collects the customer’s details and the issue, and tells them when they’ll hear back. Only promise a follow-up time if someone is actually scheduled to work that queue. “We’ll reply when we open at 9am” is fine if it’s true; “someone will contact you shortly” at midnight on a Saturday is not.

Urgent topics need their own rule. If a customer reports a safety issue or a fraud concern after hours, the bot should give the right emergency or fraud contact, not just open a ticket.

AI voice agents need the same triggers, but the handoff itself is harder. A cold transfer drops the caller into a queue with no context. A warm transfer passes the summary to the agent’s screen before the call connects, so the agent can say “I see you’re calling about a refund on your last order” in the first breath.

Tell the caller before you move them, and don’t play a long hold message on top of a long bot conversation. If the transfer fails, the call should fall back to a callback request, not a dial tone.

Customers deserve to know whether they’re talking to software or a person, and some places require it. California’s SB 1001 is one example: it makes it unlawful to use a bot online to mislead people in California about its artificial identity in order to sell something or influence a vote, unless the bot discloses that it’s a bot. This isn’t legal advice; ask your counsel which rules apply to your customers.

In practice, disclose at the start of the conversation and again at the handoff. “You’re chatting with our virtual assistant” at the top and “You’re now talking with Ana from our team” at the handoff covers it clearly, and it makes the switch to a person feel like an upgrade.

Source: California SB 1001 (2018): Bots, disclosure

The agent’s first message decides whether the handoff worked. Agents who read the summary and pick up the thread make the bot look good. Agents who open with “How can I help you today?” undo everything the bot collected.

Bot dashboards usually report containment, the share of conversations the bot finished alone. On its own, that number rewards a bot that traps people. Track the handoff as its own funnel instead, and look at these numbers together.

Every handoff is a free lesson about your bot. Read a sample of handed-off conversations every week, grouped by trigger. Low-confidence handoffs show you missing content. Repeated-failure handoffs show intents that are confused or too broad. Frustration handoffs often show a policy that customers dislike, which no bot can fix.

Score handed-off conversations on the same quality scorecard you use for people, from the bot’s first message to the agent’s last. Then change one thing at a time, so you know which fix moved the numbers. And if you’d rather not staff the takeover yourself, a team of people can sit behind the bot and pick up every handoff.