What you will learn
- Why a semantic model is more than a report data source.
- How measures, descriptions and relationships affect AI answers.
- How row- and column-level security remain relevant.
- How to validate AI-generated insights before sharing them.
The semantic model is the business contract
A warehouse may contain accurate data while remaining difficult to interpret. The semantic model names entities in business language, connects them through explicit relationships and defines measures such as revenue, margin or active customer once. Reports, analysts and AI experiences can then reason from the same definitions.
Five foundations for AI-ready analytics
Use understandable table and column names with useful descriptions.
Centralize business calculations instead of recreating them per prompt.
Model grain, cardinality and filter direction intentionally.
Test row- and column-level access for the requesting identity.
Metadata is part of the user experience
Descriptions, synonyms, formats and business examples reduce ambiguity. “Sales” may mean booked, billed or collected value. If the model does not distinguish those meanings, an AI answer can be confident and still be wrong.
Validation workflow
- Restate the business question and date/filter context.
- Identify the measure and dimensions used.
- Compare the result with a trusted query or report.
- Check missing data, unusual filters and aggregation grain.
- Confirm the user's security context.
- Explain uncertainty before recommending an action.
Portfolio project
Build a small star schema and semantic model for a sales or service scenario. Add documented measures, synonyms, RLS, an AI question set and a validation report showing correct answers, ambiguous questions and failures. Include the decision memo the analysis would support.
Primary sources
- Microsoft Fabric data agents and Power BI
- Semantic model best practices for AI
- Power BI agentic development
Microsoft feature availability varies by tenant, region, capacity and release state. Validate current product documentation before implementation.