When working with Copilot in Power BI, many users are surprised to see different answers returned for the same question. In this training session, Justin Vogel from Pragmatic Works explains why this happens and how data professionals can eliminate ambiguity by preparing their models correctly for AI. The key concept introduced is the use of Verified Answers, a feature designed to make Copilot responses consistent, predictable, and aligned with defined business logic.
Justin begins by demonstrating two nearly identical Power BI reports. Both reports use the same data and the same DAX calculations. However, Copilot behaves very differently in each one. The difference is not randomness or error. It is interpretation.
In an unprepared report, Copilot relies heavily on column names, measure names, and general wording. When users ask questions like “What are total sales?” or “What is our top performing product?”, Copilot attempts to infer meaning. That inference can vary depending on context, wording, or available metadata.
Copilot is not wrong in these scenarios. As Justin explains, it is doing the best it can with the information provided. The problem is that inference introduces inconsistency.
To remove ambiguity, Justin introduces the concept of Prep Data for AI in Power BI. One of the most powerful components of this process is Verified Answers.
Verified Answers live in the semantic model within the Power BI service. While the DAX logic that drives accurate results is usually created in Power BI Desktop, Verified Answers are configured after the report is published to a paid Fabric-enabled workspace. This feature does not work in trial environments.
Once inside the semantic model, Power BI guides users back to the report itself. Justin demonstrates how to define a verified answer by right-clicking on a visual in the report.
In the example, a visual shows the highest-selling product, which is calculated using an existing DAX measure. Instead of allowing Copilot to guess what “best-selling product” means, Justin explicitly defines that visual as the answer.
Power BI automatically suggests related questions, such as:
These suggested phrases can be added to the verified answer. Once applied, any question that closely matches these phrases will return the predefined visual as the response.
Justin makes an important distinction between measure descriptions and verified answers.
In Justin’s words, descriptions improve probability, while verified answers create certainty.
After republishing the report, Justin revisits the Copilot pane and asks the same question: “What is our bestselling product?” This time, Copilot immediately returns the exact visual that was defined as the verified answer. The Copilot pane also confirms that the question matched a verified answer.
In a fully prepared model, even questions like “What are total sales?” return precise values rather than ambiguous explanations. Copilot no longer needs to guess because the business logic has already been defined.
Justin concludes with an important message for data professionals. Preparing data for AI is not about making Copilot smarter. Copilot is already powerful. The goal is to make the model clearer.
By understanding how Copilot reads and interacts with a semantic model, users gain control over that amplification. Verified Answers ensure consistency, trust, and confidence in AI-driven insights.
For more guidance on Copilot, Fabric, and AI in Power BI, Pragmatic Works offers additional videos, on-demand learning, and consulting services to help organizations succeed.
``