Most AI support projects spend months on the conversations the agent can finish and an afternoon on the ones it can't. Customers notice. A clumsy handoff undoes the goodwill of every quick answer that came before it, because it is the moment the customer is already a little frustrated.
We have reviewed thousands of handoffs across Parley customers. The good ones share four habits.
1. Hand over early, and say why
An agent that keeps trying after it should have stopped is worse than no agent. Write the handoff rules first: a mention of a symptom, a complaint about a previous complaint, a refund above a limit, any sign of distress. When one fires, the agent says so plainly: "This needs one of our team. I'm passing you over now with everything you've told me."
2. Pass the case, not the transcript
A person picking up a conversation should not have to read forty messages. The handoff note should fit in a glance:
- What the customer wants, in one line
- What the agent already checked, with links to the order, booking or account
- Why it handed over
- What the customer has been told will happen next
In Parley this note is generated from the trace, so it lists the actions the agent actually ran, not what it thinks it did.
3. Keep the channel
If the customer started on WhatsApp, the person replies on WhatsApp. Moving someone from chat to email "for security" or to a phone queue for convenience feels like being sent to the back of the line. If the person needs to verify identity, do it inside the same conversation.
4. Close the loop with the agent
Every handoff is a lesson. The best teams review a sample each week and sort them into three piles: handoffs that were right, handoffs that were too early, and handoffs that were too late. The middle pile becomes new playbook steps. The last pile becomes new handoff rules.
Customers forgive a handoff. They don't forgive being made to repeat themselves.
What to measure
Track the time from handoff to first human reply, the share of handed-over customers who had to repeat information, and the satisfaction score for handed-over conversations on their own. If that last number is lower than for fully human conversations, the handoff itself is the problem.



