Hand Off to a Human

This is the last step of Stage 2. You set up escalation: the agent hands a conversation to a human supervisor when it detects a situation it should not handle alone, such as a decision outside policy, a request for an exception, or a safety emergency.

In the CMS, escalation settings live under Conversation → Escalation in your agent's sidebar.

If the organization has Privacy Mode on, a supervisor handling an escalation cannot see the conversation content or get a real AI summary of it. See Receive and Manage Escalations.

Before you start

• Your agent is published and reachable by users: see Preview and Publish Changes and at least one channel, such as Embed Chat on Your Website.
• You have Owner or Contributor access to the agent.
• At least one person who can act as supervisor is a user of your organization. See Manage Who Can Use and Build.

Pages in this step

This order builds each part on the previous one: rule quality first, human handling second, outbound integration last.

Why escalation matters

Without escalation, an agent can keep answering in loops, miss urgency, or fail to route sensitive requests to the right owner. Escalation gives you:

1. User protection: people with an urgent or unusual request reach a real person quickly.
2. Operational control: supervisors claim and complete cases in a clear workflow.
3. Risk reduction: sensitive situations go to humans instead of forced AI answers.
4. System integration: webhook events can notify external tools.

How escalation works

1. The agent detects a trigger condition from your escalation rules.
2. An escalation record is created with urgency, reason, topic, and area. The agent asks the user for a name and phone number if it does not have them yet.
3. The supervisor workspace updates in real time.
4. A supervisor claims the escalation and reviews the case context, the AI summary, and the conversation.
5. The supervisor writes a resolution note and completes it. An agent Owner or Contributor can also release a claimed case back to Pending.
6. An optional webhook event is sent to external systems.

The tutorials in this step use one running example: Atlas, Kirana Group's operations assistant. An employee needs to approve a USD 1.2 million project and asks for an exception to the approval limit, because the Regional Director is on leave and the vendor deadline is tomorrow. SOP-FIN-007 does not allow that, so Atlas must not decide. The expected result is that Atlas escalates to the Operations duty supervisor in real time. In the test, Rina Wijaya, Kirana's operations lead, plays both parts: she sends the request from the Preview chat, then handles it as the supervisor on duty.

Prepare before you configure

A short checklist makes your first test meaningful:

1. Trigger phrases and boundaries: list examples that should escalate (decisions the SOPs do not cover, exceptions to approval limits, safety emergencies on site) and examples that should stay with the agent (ordinary approval-chain questions the SOP answers).
2. Urgency levels: decide what counts as low, medium, high, and critical urgency, and the expected response time for each.
3. Supervisor coverage: make sure at least one supervisor can claim cases during your operating hours.
4. Labels: standardise reason, topic, and area labels so filtering and reporting stay consistent.
5. Test scripts: prepare 2 to 3 conversations that should escalate and 2 that should not, and decide who plays supervisor during testing.
6. Webhook endpoint (if you plan to use one): have a test endpoint ready to capture payloads.

Next steps

• Start with Configure Escalation.
• When escalation works, Stage 3 begins with Search and Read the Web.