Microsoft Fabric to lose auto-generated semantic models

Microsoft Fabric users will soon face more work to set up analytics workflows for new datasets, as Microsoft is retiring a feature that automatically creates semantic models on enterprise data.

Semantic models are structured representations of data that that add meaning and context to the raw information held in Fabric. When creating Fabric assets such as warehouses, lakehouses, or SQL databases, users can make their own semantic model — or use an automatically generated Default Semantic Model.

The default option, though, will soon go away by the end of this year, hurting enterprises using them for rapid prototyping or validating data structures, and any Default Semantic Models already created will become regular semantic models, that enterprises must explicitly manage and maintain for themselves, rather than relying on Microsoft’s automated processes.

Microsoft said that retiring the Default Semantic Models is part of its broader push to tighten up data governance across its enterprise offerings.

Microsoft’s strategy to emphasize accountability

Moor Insights and Strategy principal analyst Robert Kramer sees the move as a way for Microsoft to emphasize the importance of accountability in how teams build and manage data models, especially as companies scale up their AI and analytics initiatives.

“This can help organize workspaces and make reporting more transparent. Teams will have better audit trails, clearer lineage, and more accurate usage data,” Kramer said.

This change introduces an additional step for enterprise users, requiring manual creation of semantic models that potentially will slow down rapid prototyping workflows, he said.

“It’s advisable to use Microsoft’s migration script early to avoid any unexpected issues,” Kramer added.

AWS and Google are walking the same path

Microsoft is not the only hyperscaler that is betting on enterprises adopting a well-structured semantic layer instead of one auto-generated on the fly.

Google’s Looker requires users to define their own models using LookML, and more recently the hyperscaler has added Gemini to it to enable AI-generated model suggestions and chat-based analytics to accelerate insights generation.

“AWS takes a different route with QuickSight Topics — machine learning can suggest fields, but authors still curate the model. Both are pushing toward the same end goal: curated, explainable models that work well with AI and keep data governance in check,” Kramer said.

What’s changing and how will the change take effect?

After the retirement of the Default Semantic Models later this year, the process of creating a warehouse, lakehouse, SQL database, or a mirrored database in Fabric will no longer automatically generate a semantic model with the same name, Microsoft said.

At the same time, the Reporting Tab inside Fabric will no longer include options such as ‘New Report,’ ‘Manage default semantic model,’ or ‘Automatically update semantic model,’ which were previously tied to these auto-generated models.

The deprecation of the models also removes features like model layouts and context menu shortcuts that supported quick report creation using these models, Microsoft said. Instead, users will have to explicitly create and manage semantic models through a new entry point under the Home tab.

Existing Default Semantic Models will also be affected. The models will be decoupled from their parent data assets and converted into standalone semantic models, which users will need to manage manually, Microsoft said.

The changes will take place by December 2025, Microsoft said.

Clean up models now to avoid headaches later

To prepare for the changes, Kramer advised, “Enterprises can start by using the Fabric Admin APIs to find all existing Default Semantic Models and tagging them as keep, merge, or retire.”

For the models they choose to keep, he said, “They need to rebuild them as explicit semantic models using Power BI Project (PBIP) or Tabular Model Definition Language (TMDL) so that they can be versioned and governed properly.” The PBIP file format stores Power BI reports and semantic models in a way that supports version control and collaboration, while TMDL is the modeling language used within PBIP files to define and manage the structure and logic of the semantic model.

Enterprises can also add metadata to their models in Microsoft Purview, a suite for managing data governance, to improve discoverability and governance, Kramer suggested.

“Enterprises also have to ensure that analysts are trained on star schema basics so future models are clean, fast, and compliant. Teams that take the time now to clean things up and train properly will avoid headaches later with messy models and performance issues,” he said.

More Microsoft news:

Microsoft C++ static analysis tool bolsters warning suppressions

Microsoft SharePoint zero-day breach hits on-prem servers

Microsoft will stop using Chinese workers on US DoD systems