AI Escalation to Human Agents

One of the most important aspects of the AI agent is knowing when to step back and let a human take over.

When does the AI escalate?

The AI escalates to a human agent when:

• It cannot confidently answer a question based on the knowledge base
• The customer explicitly asks to speak to a human
• The conversation involves a sensitive or complex issue
• The AI detects frustration or urgency in the customer's messages

How escalation works

When the AI escalates, the full conversation history is preserved. Your agent picks up exactly where the AI left off — the customer does not need to repeat themselves. The AI can also add an internal note summarizing the conversation context to help the agent get up to speed quickly.

The escalate tool

Internally, the AI uses a single "escalate" tool to hand off a conversation to a human or team. The tool captures a reason for the handoff and routes to the team or agent configured in your AI agent settings. For anonymous widget visitors, the AI is also instructed to ask for an email before escalating, so you can follow up even if no operator is available.

When no operator is available

If the AI escalates and no operator is online (or your escalation strategy is set to the unassigned queue), the conversation is moved into the unassigned queue and flagged as awaiting a human. Two things then happen automatically.

• Operator digest — pending escalations are grouped per operator and delivered as a single digest email, ordered by priority, so a burst of escalations does not flood inboxes. See "Notifications & Alerts" for how to tune cadence, priority threshold, and quiet hours.
• Visitor follow-up — for anonymous widget visitors, the AI asks for an email and name inline before escalating. Once submitted, the contact is saved even if the visitor closes the tab. When an operator eventually replies, Savena emails the visitor the unread reply so they can pick the conversation back up. See "Following up with Visitors Who Leave the Chat" for details.