Add human escalation to your customer assistant

Teach your assistant when to hand off to a human — and wire it up to your ticketing system so the support team gets the question with full conversation context already attached.

Overview

Add human escalation to the customer assistant you built in Setup. The agent recognizes when it can't help and creates a support ticket with the full conversation as context. A safety-net button at the end lets customers reach a human at any time.

Prerequisites

You completed the Customer Assistant setup tutorial — a working sub agent connected to Inkeep Unified Search.
Your ticketing system is connected as an MCP server, exposing a create_support_ticket tool. We'll treat the MCP as a black box in this guide — only the tool's input shape matters.

Why escalation works this way

Why intent-driven is the primary path

A static "Contact support" form makes the customer leave the chat and re-explain their issue from scratch. The agent already knows the question, what it tried, and where it got stuck — so it can create the ticket with all that context already attached. The support rep opens the ticket and sees the full story.

What your ticketing MCP needs to expose

Throughout this guide, we'll assume your ticketing MCP exposes a tool with this shape:

Tool name: create_support_ticket

Inputs:

{
  "subject": "Short, 1-sentence issue summary",
  "description": "Full conversation context: what the customer asked, what the agent tried, what's still blocking. Quoted user messages where useful.",
  "customer_email": "where the support team should follow up"
}

Output:

{
  "ticket_id": "TCK-12345",
  "status": "new",
  "requester_email": "customer@example.com"
}

If your real MCP has different field names, adjust the prompt and tool call accordingly. The pattern is the same.

Attach your ticketing MCP to the agent

Make sure your ticketing MCP is registered in MCP Servers with valid credentials (same flow as Inkeep Unified Search in Setup).

Go to your agent canvas: Agents → Customer assistant.

Drag and drop an MCP block from the top left toolbar onto the canvas. Select your ticketing MCP and connect it to your sub agent.

Click Save Changes in the top right corner.

Update the prompt to escalate with intent

The agent now has the ability to create tickets, but it needs to be told when to escalate and what to put in the ticket. Without explicit instructions, the agent will either escalate too eagerly or not at all.

Add these sections to your existing sub agent prompt:

When to escalate (call `create_support_ticket`):
- The user explicitly asks for a human: "talk to a human", "speak to support", "I need help"
- You searched the knowledge base and found nothing relevant
- The user repeats the same question after you've already answered
- The user expresses frustration or sounds stuck

How to escalate:
1. If you don't have the user's email (no `x-user-email` header), ask once:
   "I'll get this to our support team — what email should they follow up on?"
   Wait for the user's reply before continuing.
2. Once you have the email, call `create_support_ticket` with:
   - `subject`: 1-sentence summary
   - `description`: full conversation context. Quote the user's question; note what you searched, what you found, what's still blocking.
   - `customer_email`: the email you collected (or from `x-user-email`)
3. After the tool returns, confirm:
   "I've opened a support ticket — our team will follow up shortly."

Tip

The bullet about the user repeating the same question after you've already answered matters. Without it, the agent keeps re-phrasing the same answer instead of recognizing the loop and escalating.

Click Save Changes in the top right corner.

Test the intent flow

Click Try it in the top right corner to open the chat interface.

Try each escalation trigger to confirm the agent handles them:

Explicit ask: "I need to talk to a human."
No docs match: Ask something your knowledge base wouldn't cover (e.g., "Can you refund my purchase from last March?").
Repeat question: Ask a real product question, then ask the same thing again after the answer.

In each case, the agent should ask for your email (if not present), then call create_support_ticket, then confirm with the follow-up message.

Add the always-on escalation button

The intent flow handles most cases, but the customer should always be able to reach a human in one click. The Inkeep widget supports custom triggers — you can add a button anywhere in your UI that submits a pre-defined message to the agent.

The pattern, in pseudo-code:

// Inside your chat component
const chatRef = useRef<AIChatFunctions | null>(null);

<>
  <InkeepEmbeddedChat
    aiChatSettings={{
      baseUrl: "...",
      appId: "...",
      chatFunctionsRef: chatRef,
    }}
  />
  <button onClick={() => chatRef.current?.submitMessage("I need to talk to a human")}>
    Talk to support
  </button>
</>

When the button is clicked, the agent receives "I need to talk to a human" — which is one of the explicit asks in the prompt — and runs the same escalation flow. No separate ticketing path needed.

Note

The exact API depends on which Inkeep widget you're using (Embedded Chat, Chat Button, Side Bar Chat). See the Chat Components Overview for the full reference.

Add human escalation to your customer assistant

On this page