With apologies to Dr. Dolittle: 

If you could talk to your data, just imagine it 

Chatting to a lakehouse in SQL 

Imagine speaking to a semantic model, whispering to a warehouse 

What a neat achievement that would be. 

Well, that pretty much sums up Microsoft’s new Fabric Data Agent! This new agent allows you to chat with your data in natural language and translates your request to the data store’s “native” language: SQL for lakehouses and warehouses, KQL for eventhouses, and DAX for semantic models. 

Of course, making sure you get the right translation and correct results is not quite as easy as it was for Dr. Dolittle, so we will provide some guidance so you can help Data Agent provide the best answers. 

Note: As of August 2025, SQL database in Fabric is in Public Preview, so there may be changes before it becomes Generally Available. 

Understanding Fabric Data Agent

Let’s dig deeper into the world of Fabric Data Agent, a tool designed to build conversational Q&A systems using generative AI. Its main job is to turn your natural language questions into SQL, DAX, or KQL queries, making it easy to access data from lakehouses, warehouses, semantic models, and eventhouses. 

Key Features 

Multi-Source Integration: You can use up to five sources per data agent. For example, you can set up a Data Agent to answer questions using content from two lakehouses, one warehouse, and two semantic models. 

Natural Language Processing: The agent translates your everyday language into data queries, making it user-friendly. 

Versatile Access: You can access the agent directly, through Power BI Copilot, or integrate it with Azure AI Foundry. 

Configuring the Data Agent

Setting up the Fabric Data Agent involves a few steps: 

Data Agent Instructions: These define the overall goal, data sources, key terms, and response guidelines. They help tailor the agent’s responses to meet your needs and ensure accurate query handling. 

Data Source Instructions: These provide detailed guidance on tables, columns, relationships, and query logic needed to tackle common or complex questions. This boosts the agent’s accuracy. 

Data Source Example Queries: Known as few-shot examples, these improve the quality of generated queries by giving the agent reference points. This helps in forming more accurate and context-aware responses. 

For lakehouses, warehouses, and eventhouses, you'll configure these in the Fabric service. But if you're using a semantic model, for the Data Source Instructions and Example Queries you'll instead use AI Prep in Power BI Desktop. 

AI Prep in Power BI Desktop 

The “Prep Data for AI” feature in Power BI Desktop is key for ensuring high-quality AI responses if your source is a semantic model. It includes: 

AI Data Schema: Select the minimum required tables and columns for AI use, ensuring streamlined data processing. 

Verified Answers: Links questions to visuals for reliable references. Supported in Import and DirectQuery models, but not available in DirectLake. 

AI Instructions: Offer context and clarify business terms, guiding the AI in providing accurate responses. 

Testing and Publishing  

For semantic models, test in Power BI Desktop to refine and retest for optimal results. For other data stores, testing is done in the Fabric service.  

As you test you can review the generated queries and refine instructions or provide examples as needed if you don’t get expected results.  

Once testing is complete, publish the data agent and share it with other users.  

Note: Ensure users have access to underlying data sources in addition to the Fabric Data Agent. 

Integration with Azure AI Foundry 

The Fabric Data Agent can be integrated with Azure AI Foundry, becoming a knowledge source for an Azure AI Agent. This involves securely generating queries over data sources, processing the data in Fabric, and combining the results with Azure AI Foundry’s logic for comprehensive responses. 

Process Steps: 

• Use the identity of the end user to generate secure queries. 
• Invoke Fabric to fetch and process the data. 
• Combine results for comprehensive responses. 

Best Practices 

Some best practices are: 

• Use a semantic model as the data source for the agent. The semantic model incorporates the relationships between tables as well as your calculated measures. The AI Prep features in Power BI Desktop let you further enrich the model to make it work well for the Data Agent. 
• Provide detailed instructions for both the overall agent and the data source. The agent instructions provide information about tone and style, general knowledge, and help direct to the right data source if there are multiple sources in your agent. Data source instructions, or AI Instructions in Power BI Desktop, provide information on what tables and columns to use and how tables are connected. 
• When sharing the Data Agent with users, make sure to also provide access to the underlying data sources.  

Refer to Best practices for configuring your data agent - Microsoft Fabric | Microsoft Learn for more information. 

Current Limitations  

As the Fabric Data Agent evolves, new features and enhancements are expected to further improve its capabilities and user experience. There are a few limitations currently: 

• Currently, only text results are available, and there is no option to visualize the data in a graph.
• The responses can be a bit slow, for example 10 to 20 seconds typically, and sometimes up to 30 seconds. Of course, everything is relative, and if the alternative is sending a request to IT and having someone write a SQL statement or add a page to a Power BI report, 30 seconds is pretty good. 
• There is no way to review questions asked by users. Power BI has a similar natural language feature called Q&A that includes tooling to review questions, so hopefully this will be added to Data Agent as well. 

You can find planned features in the Data Science section of the Fabric Roadmap at Microsoft Fabric Roadmap.  

Conclusion 

The Fabric Data Agent can help break the logjam of waiting for an expert to create a new report every time a user asks another question. Using natural language can help democratize data analysis throughout an organization. However, preparation is critical to ensure you get the results you are looking for. 

As a Microsoft Partner with 10 Advanced Specializations, a Featured Fabric Partner, and a Fabric Databases Featured Partner, Spyglass MTG can assist with any aspect of Fabric, as well as other Microsoft technologies including AI, Azure, and PowerPlatform. We’re here to help, so if you have any questions, feel free to reach out to kfeit@spyglassmtg.com or reach out to our info email at info@spyglassmtg.com.