Skip to content

Human handoff

When the AI steps aside, how the ticket reaches an agent, and why the bot then stays quiet.

2 min readUpdated August 2026
On this page

What triggers a handoff

TriggerNotes
Money or judgement intentRefunds, billing, cancellations — always
Explicit request"Talk to a human", "is anyone there"
Low confidenceBelow your reply threshold
Intent set to EscalateYour per-intent override
AI quota exhaustedDegrades to humans rather than failing

What the customer sees

The AI sends a short holding message acknowledging the request and saying a person is picking it up. It does not pretend to be a human, and it does not attempt a partial answer to a question it has decided it should not answer.

Then it stops. This is important: once a conversation is handed off, the AI will not reply over the top of the agent who now owns it. Nothing is more damaging to a support conversation than a bot interrupting a human mid-sentence with a generic answer.

What the agent sees

The escalated ticket carries:

  • The full customer conversation
  • The detected intent and confidence score
  • The handoff reason — why the AI stepped aside
  • An AI summary of what the customer wants

The summary is the point. An agent picking up an escalation should not have to read a twelve-message thread to work out what is being asked.

Routing an escalation

Escalated tickets route by intent — billing questions to the Billing team, everything else to Support — and are then assigned by round robin like any other ticket. Their SLA clock has been running since the customer's first message, not since the handoff.

See Assignment and routing.

Handing back

There is no automatic hand-back. Once a human owns a conversation, they own it until they resolve it. If a customer opens a new conversation later, the AI handles it from the start as normal.

Tuning the handoff rate

A handoff rate above roughly half usually means thin knowledge rather than a bad threshold. Read the handoff reasons: if most say low confidence on questions you have documented, the articles need to be more direct. If most are explicit human requests, look at whether your holding message is inviting them.

Was this page helpful?