Fabric IQ and Agentic Authoring Are Now GA: What It Means for Your Power BI Semantic Model

The most important Fabric news of this season is not a new component. It is that Microsoft stopped treating the semantic model as something only analysts read.

At FabCon Europe 2026, Fabric IQ reached general availability in Microsoft 365 Copilot Chat and Copilot Cowork. Business users can now ask questions in Copilot and get answers grounded in your governed Power BI semantic models and reports — on by default for Fabric and Power BI customers, with no additional AI token cost. The same event moved PBIP (Power BI Projects) to GA, shipped agentic authoring tools, and put a new set of DAX functions into preview.

If your team already owns a semantic model, this changes the job of the analyst. If it does not, it changes the reason to build one.

What Fabric IQ actually does

The promise of "ask your data" has been around for years and always failed at the same place: the model. Large language models do not know that Receita is net of tax, or that a cancelled order should not count, or which of the four "revenue" columns is the one the CFO uses.

Fabric IQ solves that by making the semantic model the source of truth. Copilot answers from the measures, relationships and RLS you already built — the definitions, not a guess over raw tables. That is the whole value: the numbers the business gets in a chat are the same numbers the Power BI report shows, because they are the same model.

It also opens the model to things outside Power BI. An MCP server (GA) lets external agents query Fabric IQ, so a custom assistant or a third-party tool can hit the same governed layer. And IQ sharing (preview) lets organizations share governed tables, files, agent instructions and ontologies across tenants without copying data.

Why governance is now the bottleneck

Here is the practical consequence teams should not miss: the moment the semantic model feeds a Copilot answer, every messy corner of that model becomes user-facing.

• A measure that is wrong is now wrong in a chat. Someone asks a question in Copilot, gets a confident number, and forwards it. There is no report title to blame.
• Ambiguous names get blindly picked. If you have Sales, Sales USD and Sales (legacy), the model needs to make the right one obvious, or Copilot will not.
• Read access changes. Fabric IQ grounds answers for business users, so who can see what is no longer controlled only by who opens which report. Model owners can now block read-only users from querying a model via Copilot and MCP — a control that did not exist before, and one you should decide on deliberately instead of leaving at the default.

The teams that win this shift are not the ones with the fanciest model. They are the ones whose measures have descriptions, whose columns are named like a business glossary, and whose RLS already reflects reality.

The authoring side moved too

The second half of the announcement is about who builds. Agentic authoring tools went GA: an authoring MCP server, a Desktop Bridge, and skills for Fabric. In practice, a coding agent or Copilot can now create, test and deploy Fabric items — Dataflow Gen2 through the Data Factory MCP, warehouses through the Warehouse MCP server, activators and eventhouses through their own remote MCP servers.

That does not replace a modeler. It replaces the boilerplate: scaffolding a notebook, generating the first version of a pipeline, writing the T-SQL skeleton for a warehouse table. The judgment — which grain, which relationship, which measure belongs in the model — still comes from a person who understands the business.

Two supporting changes are worth noting. PBIP is now GA and PBIR is the default report format, so if you have scripts that parse the old report file format, they need attention. And the language itself grew: TEXTCONTAINS() and TEXTSIMILARITY() arrived in preview, backed by full-text search and string indexing — useful for the endless "find the customer whose name is roughly this" requests — while APPROXIMATEDISTINCTCOUNT now works for Import and Direct Lake at roughly 1.6% standard error for faster counts on large tables.

What is real today and what is still preview

Worth being precise, because the summary posts blur the line:

• Generally available now: Fabric IQ in Microsoft 365 Copilot Chat and Copilot Cowork, the Fabric IQ MCP server, PBIP and PBIR as the default report format, the agentic authoring tools (authoring MCP, Desktop Bridge, skills for Fabric), and native DiskANN vector indexing in T-SQL.
• Preview: ontologies (GA promised in the coming months), IQ sharing, the new Power Query editor, TEXTCONTAINS() and TEXTSIMILARITY(), APPROXIMATEDISTINCTCOUNT on Import and Direct Lake, and the natural-language app-building experience built on Fabric Apps.
• Policy change, not a feature: from October, Power BI Desktop versions from March 2026 or earlier can no longer save to OneDrive or SharePoint. If anyone on the team is still on an old build, that decision was made for them.

How to get ready, in order

We would not start with Copilot. We would start with the model, because that is what Copilot reads.

1. Audit the measures that matter. Every measure a business user would ask about needs a description and an unambiguous name. If two measures could answer the same question, one of them should probably be hidden.
2. Decide the Copilot access policy now. Model owners can block read-only users from querying through Copilot and MCP. Make that a deliberate call per model, not a default nobody noticed.
3. Clean the names before the ontology. The entities of a Fabric IQ ontology map to business vocabulary. That vocabulary is easier to define while you are fixing column names, not after.
4. Update the client tooling. Old Desktop builds cannot save to OneDrive or SharePoint from October. Patch or expect failure.
5. Then turn on Fabric IQ. It is on by default for Fabric and Power BI customers, at no extra AI token cost, which means it is already live in the tenant whether the team planned for it or not.

The point

The semantic model stopped being an internal artifact and became the interface between your data and everyone who asks questions about it. Fabric IQ makes that official.

That is good news for teams who invested in the model and uncomfortable news for teams who treat it as a byproduct of report building. The work that was always the real work — clear definitions, clean relationships, honest security — is now the thing users experience directly.

If your organization has Fabric capacity and no one is sure the semantic model is ready to answer questions, that readiness review is exactly what we do — see our Power BI consulting page.