Skip to main content
Human handoff moves a conversation out of autonomous AI processing and into the shared inbox. It is a state transition, not simply a different AI response.

​
Handoff triggers

  • The model calls escalate_to_human because it cannot safely complete the request.
  • AI is disabled for the widget or organization and fallback is enabled.
  • Subscription policy prevents AI processing and fallback is enabled.
  • The selected provider fails and automatic fallback is allowed.
  • An operator routes or assigns the conversation manually.

​
Flow

1

Record the reason

The agent calls the gateway with organization and conversation context plus a concise reason.
2

Change the conversation gate

The gateway marks the conversation escalated/assigned or unassigned for the human queue and emits notifications/events.
3

Suppress future AI replies

Delayed AI jobs re-check the conversation gate and stop when a human owns the conversation.
4

Operator takes over

An agent opens the inbox, assigns or routes the conversation, uses templates or assist features, and updates status when complete.
Widget configuration exposes fallbackToAgent, allowing an organization to disable escalation where no human team is available. When fallback is off, provider or subscription failures return only the configured safe behavior.
Agent-assist actions—suggest reply, generate note, and draft assist—can support a human without returning the conversation to autonomous mode.

​
Operational goals

Track escalation rate, reason categories, time to first human response, unassigned backlog, reassignment count, resolution time, and cases where the user never received a response.
Last modified on July 17, 2026