Optimizing multi-agent systems built with Microsoft Copilot Studio poses a significant challenge: most user interactions - simple Q&A exchanges - lack structured feedback suitable for traditional analytics. Are users clarifying their tasks? Questioning model quality? Reporting confusion? Without answers, it’s hard to decide whether to retrain models, educate users, or redesign the agent architecture.

In this session, we present a practical framework that uses LLM-based classification to characterize user intent. Each incoming user message is automatically labeled by type (clarification, feedback, additional info, etc.), forming a structured event log for analysis.

We’ll walk through:

- Using Microsoft Copilot Studio to enrich agent telemetry and route structured data to Application Insights
- Applying LLM pipelines to tag conversation intents
- Leveraging Power Automate Process Mining to detect interaction bottlenecks and user friction points
- Making actionable decisions: from training material improvements to architecture redesign

This is a highly practical, demo-driven session for:

- Power Platform developers building AI-powered solutions
- Architects designing conversational interfaces
- System analysts optimizing Copilot usage in enterprise environments

Join us to discover how structured data and process insights can turn black-box agent behavior into measurable, improvable systems.