Your Data Agent doesn't understand your business (and it's not the model's fault). You connect it to your lakehouse, ask how many active customers there are, and it answers with complete confidence. The problem is that in your company "active customer" has three definitions, and the agent picked whichever one it liked. You gave it tables and columns, not the meaning of your business.

In this session we take the same data and the same bank of business questions, and solve them with three ways of giving the agent semantics: a Power BI semantic model with descriptions and instructions, an ontology in Fabric IQ, and a relationship graph. We measure how many questions each approach gets right, starting from a baseline with no semantics, and see where each one breaks.

You'll leave with concrete criteria to decide: what type of question each approach solves, who maintains it, how much it costs to build, and what happens when the business changes a definition. Less "the agent hallucinates" and more "the agent didn't know what you did".