A conversational AI system loses containment when users abandon self-service and request a human. Containment measures how many interactions finish without human involvement, making it a practical indicator of cost reduction and business value. section 1.3.3 section 11.0
Treat containment as a user-experience measure, not just a bot score. Confusing questions, excessive turns, weak understanding, or a rigid dialogue can frustrate users into opting out even when the system is close to completing the task. section 8.1.2 section 8.3
A claim lookup that nearly succeeds
After answering several questions, a caller reaches a successful claim lookup. The assistant asks whether to read the details or fax them, and the caller requests an agent. Removing that extra decision and reading the details immediately preserves the completed self-service interaction; the fax option can follow afterward. section 8.1.2
Improve containment by:
- Finding where opt-outs occur and investigating whether users doubt the system's capability or encounter weak understanding. section 11.7
- Reducing unnecessary turns, clarifying confusing questions, shortening lengthy messages, and providing the most useful result first. section 1.3.3 section 8.1.2
- Tracking containment alongside user satisfaction, time to resolution, average handle time, and human touches rather than treating it as the only success measure. section 1.3.3
