1.4. Reflection Checkpoint

Key Takeaways

Before proceeding, ensure you can:

• Explain why LLMs can produce fluent but incorrect responses (fabrication mechanism)
• Define grounding and describe at least three types of grounding sources in Microsoft 365
• Distinguish between training (past, fixed) and grounding (present, queryable)
• Articulate the human-in-the-loop model and why it matters for professional accountability

Connecting Forward

In Phase 2, you will see how these first principles play out concretely: how Microsoft architects Copilot to keep your data private, how different types of context change what Copilot can do, and what makes a "chat with Copilot" fundamentally different from using a "Copilot agent."

Self-Check Questions

1. Your colleague argues that Copilot fabrications are rare because "the AI is trained on so much data it usually knows the answer." How would you correct this understanding?
2. You need Copilot to answer a question using your company's internal pricing data. What must you do to ensure the response is grounded in that data rather than in the model's general training?