The handoff is the part everyone gets wrong
Most AI support deployments are judged on the conversations the agent could not finish. A good handoff is a design problem, not a fallback.
Vinita Rathi
Product and conversation design
No agent resolves everything, and any vendor telling you otherwise is selling the demo rather than the deployment. Which means the quality of your automation is largely decided by what happens in the conversations it cannot finish.
The three bad handoffs
The cliff
The bot says 'let me connect you to an agent' and the conversation goes silent for forty minutes. The customer has no idea whether anything is happening. This is worse than having no bot, because the bot created an expectation and then dropped it.
The amnesiac
A human picks up and opens with 'Hi, how can I help?', after the customer has already explained the problem twice. Every repetition is a fresh reminder that they are talking to a system that does not retain anything.
The hostage
There is no way out. The customer types 'agent', 'human', 'representative', and the bot keeps offering FAQ links. People remember this one, and they tell other people about it.
What a good handoff carries
When the conversation moves to a person, that person should open the thread already knowing:
- The full transcript, including what the agent tried.
- Every tool call and its result, the order it looked up, the refund window it checked, and any call that failed.
- The identity the conversation was bound to, and what account context that resolved to.
- Why it escalated: agent classified as out of scope, tool failure, explicit customer request, or sentiment threshold. These need different responses.
- What the agent has already promised the customer, which the human is now committed to.
That last point causes more damage than any other. An agent that said 'someone will refund you today' has made a commitment. If the human disagrees with it, the customer experiences it as being lied to, not as a policy nuance.
Escalate early on the right signals
Waiting for the agent to exhaust itself is the wrong trigger. Cheaper and better signals exist:
- 1Explicit request. Someone asking for a human gets a human, immediately, with no retention attempt. Treat this as non-negotiable.
- 2Repetition. The same question asked twice means the first answer failed, whatever the agent thinks.
- 3Tool failure. If the lookup errored, the agent is now guessing. Escalate rather than let it improvise.
- 4Stakes. Money moving, an account closing, anything legal or medical, route on the topic, not on how well the conversation is going.
- 5Sentiment. Not as the primary trigger, but a sharp turn is worth acting on before the agent tries once more.
Set the expectation you can meet
'Connecting you now' is a promise about time. If your team is asleep, say so with the actual hours and a commitment for the morning. Customers are markedly tolerant of a wait they were told about, and completely intolerant of silence they were not warned about.
The same applies going back the other way. When the human is finished, be explicit about whether the automation resumes, because a customer who thinks they are still talking to a person and gets a templated reply has just been taught not to trust the channel.
- Support
- Design
- Operations
Vinita Rathi
Product and conversation design
Writes about where automation belongs in a support conversation, and where it does not.
Keep reading
- Playbooks
Decision trees don't survive contact with customers
Every scripted bot works in the demo and falls over in production. The failure is structural, not a matter of adding more branches, and the fix is narrower than replacing the whole thing with an LLM.
- Playbooks
How to move a customer support workflow from n8n to RoutMe
A step-by-step migration of a WhatsApp support workflow from n8n to RoutMe Flows: node mapping, the agent step, connectors and a checklist to cut over safely.
See it on your own data
Connect a system, drop an agent node into a flow, and watch it answer from what is actually true in your systems right now.