We are currently experiencing a generational surge in AI use cases and transformation. But will AI be able to connect directly to your raw data, perform all the necessary transformations, apply business-friendly names to fields, and add accurate logic to solutions? Does data engineering and architecture still matter?
The video embedded at the bottom of this article is a recorded version of the presentation that I gave at SQL Saturday 2025 Minnesota. In February I will be presenting an updated version of this presentation at the Power BI & Fabric Summit which is organized by Radacad. You can register for the upcoming Summit at this link: https://fabricsummit.com/product/PowerBI-Fabric-Summit-2026 . I expect these tools to evolve quite a bit over the next few months!
A well-known limitation of AI—specifically large language models (LLMs)—is that they are not fundamentally designed to perform accurate math. While newer LLMs can handle some mathematical tasks, query speeds are often slow, and the compute costs can be high. Translating the specialized context of natural language questions into precise logic also presents challenges. For example, if a business user asks, “Show me total sales for the year,” what exactly does “year” mean? Is it a calendar year, a fiscal year, or year-to-date?
Now imagine how much more complex the math becomes with a question like, “Show average sales for blue and red widgets for customers in the East, excluding store holidays.” Traditional best practices known by data professionals for decades provide a solid foundation for accurate math in AI-driven applications. These practices will continue to evolve as we design data architectures optimized for AI. Microsoft Fabric semantic models are a powerful tool for building that logic in a way that provides both context for accurate calculations and fast, efficient query performance.
If you’re a data professional with skills in dimensional modeling, query optimization, ETL/ELT, RLS/OLS—your expertise is now more crucial than ever for AI solutions that require “good math.” This video explores strategic reasons for using semantic models as the foundation for AI when querying structured data. It walks through a use case that begins with 275 million rows of raw data, demonstrates how to model the data for AI, leverages tools in Fabric semantic models to prepare the data, and then serves it to AI tools and agents using Fabric Data Agents, Power BI Copilot, Azure AI Foundry, and Microsoft 365 Copilot. GitHub repo to (no code) deploy the demo in the video to Fabric with 275M rows of real data: https://github.com/isinghrana/fabric-samples-healthcare/tree/main/analytics-bi-directlake-starschema
Here are a few helpful links:
Official Microsoft Blog article on this topic: Power BI semantic models as accelerators for AI-enabled consumption | Microsoft Power BI Blog | Microsoft Power BI
Related article on this topic from a Partner: AI in Power BI: Time to pay attention – SQLBI
Leave a Reply