Most advice about bot handoffs is about the trigger: when should the bot give up? That matters, but customers mostly judge a handoff by what happens in the next thirty seconds. Did the agent already know the problem, or did the customer have to start again? If they had to repeat themselves, the handoff failed, however clever the trigger was.
This guide works backwards from the agent who receives the conversation. Decide what that person needs, then build the trigger, the wording and the routing to deliver it.

Define the package the agent receives
Write down the handoff package before you configure anything. A good one fits on a screen and contains:
- Reason for handoff. One of a short fixed list: customer asked for a person, bot could not find an answer, topic requires a person, customer sounded frustrated, bot failed twice on the same step.
- One-line summary. What the customer wants, in plain words.
- Details already collected. Name, account email, order or plan identifier, what they tried. Anything the bot asked for should be visible, not buried in the transcript.
- The full transcript, linked or inline, with the customer's last message visible first.
- What the bot already told them. Agents need to know if the customer was already given an answer they disagree with.
- Channel and time. A phone message left at night is a different job from a live chat.
An example of a note an agent could receive:
Example textReason: no answer found in our policies. Customer says a promotional code was rejected at checkout and wants to know why. Collected: account email, the code, the screen where it failed. Told so far: the code's published terms. Customer's last message: "That does not explain it, the code is still inside the dates."
An agent can act on that in ten seconds. Compare it with a transcript dump, which makes them reconstruct the case.
Choose triggers you can explain to an agent
Triggers fall into three groups. Use all three, and keep each one simple enough to explain in a sentence.
- Customer-requested. The customer asks for a person, in any wording. Honor it immediately. Making someone argue with a bot to reach a human is the fastest way to turn a small question into a complaint.
- Topic-based. A list of subjects that always go to a person: account ownership disputes, safety concerns, complaints about an agent, anything your policy says a person must decide. Maintain this list with your team lead, not with the bot's configuration alone.
- Struggle-based. The bot has failed to answer twice, the customer repeats the same question, or they send the same message again with more punctuation. Two failed attempts is a common starting point; tune it after reading real conversations.
Do not trigger on sentiment alone without a person checking the results. Frustration detection misfires, and a false trigger sends easy cases to an expensive queue.
Tell the customer what happens next
The words the bot uses at the moment of handoff decide whether the wait feels acceptable. Say three things: that a person is taking over, what they already have, and what to expect.
Example textI'm passing this to our support team. They'll see what we've covered so far, so you won't need to repeat it. Someone will reply by email within the hours listed on our contact page.
Only promise a time you can keep. If nobody is staffed overnight, say so and offer to take a message instead of pretending a live agent is seconds away. A handoff to an empty queue is worse than no handoff.
Match the handoff to the channel
The same package works differently by channel:
- Chat, staffed hours: transfer live, with the package pinned at the top of the conversation.
- Chat, outside hours: collect the details and an email address, then create a ticket or message for the first shift, flagged as a handoff.
- Email: the "handoff" is routing. Send the thread to the right queue with the summary at the top, and tell the customer who will reply.
- Phone: transfer to a number that is actually answered, or take a message with the reason and callback details, then deliver it to the team promptly.
Whichever you use, test the out-of-hours path first. It is the one that quietly fails.
Keep the queue side honest
A good package is wasted if handoffs land in a general pile. Give handoffs their own flag or view so agents can see them as they arrive, and set a target for first reply that is at least as strict as the one for ordinary tickets, since these customers have already spent time with the bot. If you have several shifts or locations, route by topic and hours, not by whoever happens to be online, and name a fallback owner for each route so a handoff never waits on a person who is out.
Train agents on the opening line too. The customer has just been told the team can see everything, so the first human message should prove it:
Example textThanks for waiting. I can see you tried the code at checkout and it was rejected even though it's inside the dates. Let me look into that.
An opening that asks the customer to restate the problem undoes the work the package did.
Review handoffs weekly
Once a week, read ten handoffs end to end, from the customer's first message to the agent's reply. Ask whether the reason was right, whether the summary was accurate, and whether the agent had to ask anything the bot had already collected. Fix the package before you touch the triggers. Count the handoffs that should not have happened as well, because a bot that hands off everything is just a slow front door.
What not to do
Do not hide the human option or make the bot pretend to be one. Do not hand off with only "a customer needs help". And do not measure success by lowest handoff rate: a team that pushes the rate down by making people harder to reach will see it show up as repeat contacts and unhappy reviews, a pattern covered in how to reduce repetitive support tickets.
Where RyteDesk fits
RyteDesk is an AI customer support assistant in early access, built for mid-market teams. It answers from the knowledge and policies you provide on chat, email and phone, tells callers it is an AI, and hands off to your team or takes a message when it cannot help, so the person picking up sees the details already collected. You can test it in a preview before it speaks to customers. Join the waitlist if you want to try it; it is free to join.