1.2.1. Why Language Models Need Grounding
💡 First Principle: A model can only be as current and as private as its training data — which is frozen in the past and contains none of your tenant's content. Hallucination is not a bug to patch but a structural property to engineer around: the model always generates; grounding controls from what.
Three gaps make grounding non-negotiable for enterprise agents. The freshness gap: training data has a cutoff, but your price list changed yesterday. The privacy gap: the model has never seen your SharePoint, your Dataverse tables, or your ticketing system — and licensing your data into a foundation model is neither possible nor desirable. The accountability gap: enterprises need answers with citations, so a human can verify the source; free-form generation offers fluency but no provenance.
Grounded generation addresses all three at once: retrieve current, permission-checked content; generate from it; cite it. This is also why Copilot Studio's generative answers surface citations — the citation is not decoration, it is the audit trail.
⚠️ Exam Trap: "The agent gives outdated answers" scenarios are knowledge problems (stale or missing sources), not model problems. The fix is adding or updating a knowledge source — not switching models, not more instructions.
Reflection Question: Your agent confidently cites a vacation policy that was replaced last month. Walk the three gaps: which one failed, and what does that tell you about where the fix lives?
