An agent reasoning over badly defined concepts produces badly defined conclusions — only now with far more eloquence and at far greater scale. This session builds, step by step, the layer that prevents it.

We start from a deliberately dirty business domain — three definitions of "active customer", two incompatible product hierarchies, one KPI with four different calculations live in production — and take it to a model an agent can query without inventing. We'll work through the glossary with owner and version, the disambiguation rules, how it lands in Microsoft Fabric and Data Agents, and the evaluation set that proves the fix actually held rather than just looking right on paper.

The output is the business model, not the agent. Every participant leaves with the repository, the templates and a runnable test suite to point at their own model.