Inbox & Handoffs

How handoff works

When an AI assistant determines it cannot adequately help a user, it initiates a handoff to a human agent. This is a core safety and quality mechanism in comxbot.

Handoff triggers include: user explicitly requests a human, sentiment drops below threshold, assistant confidence is low, specific topics are detected (e.g. complaints, legal issues), or conversation exceeds turn limit without resolution.

When triggered, the assistant informs the user that a human will take over, the conversation enters the Inbox queue, and an assigned agent is notified.

The transition is seamless for the user. The chat continues in the same thread. The agent sees the full context and can respond directly.

If no agent is available, the assistant can be configured to: collect contact info and promise follow-up, continue attempting to help, or provide alternative resources.

Configuring handoff rules

Navigate to the assistant settings, then the Handoff Rules section. Here you define when and how handoffs occur.

Sentiment threshold: Set a score (1-5) below which automatic handoff triggers. Default is 2. The assistant detects frustration, confusion, or dissatisfaction signals.

Topic-based rules: Define keywords or topics that always trigger handoff (e.g. 'refund', 'legal', 'complaint', 'cancel subscription'). These override confidence-based logic.

Turn limit: Maximum conversation turns before forced handoff. Prevents users from being stuck in loops. Default is 15 turns.

Business hours: Configure available hours. Outside business hours, the assistant can collect details and create a ticket for follow-up rather than attempting live handoff.

Risk routing: High-risk conversations (detected via sentiment and topic analysis) are routed to senior agents. Configure risk levels and team assignments in Routing settings.

Cooldown: After a handoff is resolved, there is a configurable cooldown before the same user triggers another handoff (prevents abuse). Default is 5 minutes.

The handoff packet explained

When a conversation is handed off, the receiving agent gets a structured packet with full context:

Transcript: The complete conversation history, including all user messages and assistant responses.

Summary: An AI-generated summary of the conversation, highlighting the user's issue, what was attempted, and why handoff was triggered.

Intent: The detected user intent (e.g. 'wants refund', 'needs technical help', 'billing question').

Sentiment: Current sentiment score and trajectory (improving, declining, stable).

Citations: All knowledge sources referenced during the conversation, with links.

Suggested Next Steps: AI-recommended actions for the agent based on the conversation context.

User Info: Any identified user information (name, email, account) if available from the chat context or authentication.

This structured packet means agents never start cold. They have full context within seconds of accepting a handoff.

Managing threads and assignments

The Inbox page shows all active handoff threads. Use filters to view by status (Open, In Progress, Resolved), assignment, priority, or date.

Claiming: Agents can claim unassigned threads. Once claimed, the thread is locked to that agent (others can still view but not respond).

Reassignment: Admins can reassign threads between agents. Useful for shift changes or specialist routing.

Resolution: Mark a thread as resolved when the issue is complete. The user receives a satisfaction survey (configurable). Resolution data feeds into Analytics.

Internal notes: Agents can add internal notes visible only to team members. Useful for context-passing between shifts.

SLA tracking: Configure target response times. The inbox shows overdue threads highlighted in red. SLA metrics are tracked in Analytics.

Bulk actions: Select multiple threads to bulk-assign, bulk-resolve, or bulk-tag. Useful for managing high-volume periods.