It is no secret that the AI revolution has started. Most every modern business application has an agent, a copilot or assistant asking to help write a document or take meeting notes. Nearly every banking or telecom customer service site greets us with a chatbot. Microsoft Teams meetings are concluded with Copilot summarizing the discussion and talking points. Like Chat GTP, Copilot serves as a next-level search engine that can spit-out code snippets and summaries from blog posts and product documentation, but what’s next? When can I say “hey, Copilot, gather up data from our company line of business databases and forecast next year’s profitability by month and region.” Will AI reach that level of sophistication? Before we get into AI, let’s cover a topic that is super basic and super important in this context.
Fabric Solution Architecture and Power BI
The following diagram is a very common example of a simplified solution architecture that I see all the time. because the diagram is simplified, I’ll point out that it depicts Power BI used to simply “create reports” from the data warehouse in the gold layer. I’ll get back to that.
Unfortunately, many data architects and enterprise customers believe that this is a sufficient solution to meet their business reporting needs. In an enterprise scenario, and particularly if you plan to embrace Copilot, Power BI is not a simple tool that you just “point at data” and expect to work at scale.
On the left side of the diagram, we have a robust data engineering process to ingest data from multiple sources. In detail, this might include a sophisticated metadata-driven framework to handle incremental loads and updates. In the silver layer, notebooks and dataflows scrub and reshape data. The gold layer contains data shaped into dimensional tables; facts and dimensions. Can’t report developers just connect to the warehouse and other sources, and just create functional reports? No! This never has and probably never will be the case. Sure, with help from AI, new tooling will assist but not eliminate this important step. Other products with that promise have come and gone.
Semantic modeling is necessary, and especially for certified and business-governed report solutions, this architecture is not sufficient because it doesn’t include the crucial semantic modeling layer.
Product marketing and surface-level features in Fabric suggest that business-ready semantic models are magically created with every lakehouse, warehouse or database. Modeling for enterprise and AI-ready BI reporting will take more work and require more attention to best-practice design patterns than ever before.
Having attended several dozen meetings with Microsoft engineering teams and executive leaders recently and this week at the 2025 Fabric Community Conference, the next generation of AI enablement across the analytic reporting and data platform is becoming clear. Power BI nor any other analytic reporting tool will not magically create useful business reports using AI alone. Sorry to break the news but it’s just not going to happen, at least not in the foreseeable future.
The new Copilot integrations for Power BI provide a conversational interface for Power BI in the web portal Home page and in the web-based report designer. This is a journey that the product development teams are only a few months into. They have made tremendous progress, but it will take time for Power BI Copilot to be a seamless and reliable experience, requiring both the technology and the human user to adapt to new experiences. Although promising, we have been testing Copilot on existing models, and Copilot hasn’t really saved much work or built impressive reports on its own.
AI can help developers and savvy business users accelerate the process, but we still cannot ignore the basic science of data cleansing, preparation, and semantic modeling. In fact, disciplined semantic modeling is far more important than ever before. More than ever, BI developers must follow the rules of dimensional modeling, write efficient DAX measures and make sure that semantic models are clean, organized, use friendly naming conventions, hierarchies and display folders. This short document reinforces that these best practices are as critical is ever: Update your data model to work well with Copilot for Power BI – Power BI | Microsoft Learn
AI enablement is evolving
So, what is happening now and what should we expect to see in the next six to twelve months? Modern data tooling will continue to include agentic AI supported by large language models and learning algorithms. Every programming tool will have a coding assistant that can translate conversational prompts into program code gleaned from samples in product documentation and Internet sources.
The following solution diagram introduces a layer for governed semantic models and standardized business reports. Since general purpose chat LLM services like Copilot don’t inherently understand business rules of a semantic model, special-purpose agents can be paired with models to guide a report developer or self-service user to consume the model in the proper context. They will provide guardrails and boundaries so each model can be used correctly for its intended purpose. Copilot can be used as the standard chat interface utilizing the model-specific agent or perhaps multiple agents.
Copilot for Power BI developers is a pretty good experience because LLMs are particularly helpful for finding code examples and assisting developers write and debug code. For the Power BI user and departmental report creator, the Copilot experience on the Power BI service Home page and web-based report designer are still evolving. That evolution will be a combination of improved Copilot and Power BI features aided by developers learning how to create optimized semantic models. Until then, results will be hit or miss. Kurt Buhler documented a complete walkthrough of his experience using Copilot with Power BI in this set of Data Goblins blog posts: Myths, Magic, and Copilot for Power BI — DATA GOBLINS
For now, think of Copilot as the next generation of the Power BI Q&A feature, which has been around since the initial release of the Power BI service. If you haven’t used Q&A in Power BI, think of it as a dumb version of Copilot that allowed users to “talk” to a model and visualize results without the guidance of an LLM. Outside of demos, Q&A didn’t get a lot of uptake but it does effectively translate a natural language prompt into a report visual. Copilot aims to do the same thing with the guidance of a learning model. Like Q&A, enhancing models with synonyms helps expand the vocabulary for user-prompted Copilot conversations. Other features like verified answers will follow, along with schema selection and special-purpose agents with pre-defined business rules.
Prepare Power BI today to work with AI in the future
I would love to tell you that there are new best practices that will guarantee seamless AI integration. There are some new patterns that will help but the fact is that you must cover the basics first.
- Start by separating one-off reports, that have internal semantic models/dataset, from those that should utilize a common semantic model.
- Consolidate semantic models that can serve reports and report users that use the same common set of tables and metrics.
- Follow the star schema design pattern. Complex models may at times require some snowflake tables (dimensions related to dimensions) or duplication of summary fact tables and low-level details, but there are well-established patterns for meeting these requirements. After building POCs and work-arounds, use iterative design to simplify the model. Complicated, over-engineered models will not work with Copilot.
- Follow the published guidelines to optimize models for Q&A… add synonyms to provide alternate phrases for common business attributes.
- Create explicit measures and remove or hide all columns not needed for reporting.
The Power BI Copilot story will evolve quickly, but these are steps you can take to get ready and prepare your business for the future of AI-enabled reporting.
