Microsoft Is Forcing Power BI Semantic Models Into One Governed Layer

Somewhere in your tenant there's a semantic model with a half-finished date table and three measures named test. Nobody's opened it since twenty seventeen. As of this month, that model can talk.

Talk to who?

To everyone. Fabric IQ is now generally available inside Microsoft Copilot Chat and Copilot Cowork. Which means the definitions sitting in your Tabular models... your revenue measure, your customer dimension... are the thing answering questions in Microsoft three sixty-five.

So a measure some analyst wrote years ago, then left the company, is now feeding a chatbot the whole business uses.

And there's a line buried in a Microsoft blog post from September twenty-ninth that frames the whole thing. Quote: the semantic layer is an organization's most strategic asset.

Strategic asset. That's vendor-speak for we're about to charge you for it.

Partly. Hold that thought, because by the end I want to show you the one word in Microsoft's own migration docs that tells you where the money goes.

One word.

One word. First, the problem. Most large organizations don't have one semantic model. They have forty. Finance has theirs, supply chain has theirs, somebody in a regional office built one in Analysis Services and moved on.

And every one of them defines gross margin slightly differently.

Survivable when a human reads the report and knows which one to trust. Not survivable when a language model reads all forty and picks one.

Okay, so what's Microsoft actually DOING about it, as opposed to blogging about it?

Three moves. First: metadata moved to the browser. The July twenty twenty-six release shipped TMDL view on the web. Tabular Model Definition Language, in a code editor, in the service. Microsoft's words: immediate, code-level visibility into all semantic model metadata.

No more downloading a pbix file and praying.

That's the point. The model becomes text. Text goes into source control. Source control means diffs, review, approvals. They also added measure descriptions as triple-slash comments inline, so the documentation lives in the same file as the logic.

Which only matters if anyone writes the descriptions.

It matters more now, because the AI reads them. A vague description used to embarrass you in a design review. Now it produces a wrong number in someone's Copilot chat.

Huh. So documentation stops being hygiene and becomes runtime behavior.

Second move landed days ago. The September twenty twenty-six feature summary gives semantic model editors control over how Copilot can use their data. If you have write permissions, you can block read-only users from reaching that model through Copilot... and through Microsoft's remote MCP server.

Wait, why would you ever block people from asking questions?

Remember the model with three measures named test? Today, a read-only user has to go hunting for it. Tomorrow, Copilot surfaces it unprompted. With total confidence.

So the governance question flips. It's not who can see the report. It's which of our models deserve to speak.

Take that sentence into your next steering meeting. Now, third move. The quiet one. Microsoft's migration guidance page. Quote: currently there aren't any plans to deprecate Azure Analysis Services.

Currently.

That's the word. The very next sentence says there's a priority to focus investment on Power BI in Fabric for enterprise data modeling, and that the additional value of Power BI will increase over time. No deprecation date. Confirmed. But a widening feature gap, stated openly by the vendor.

That's not a deadline. That's a drift you can ignore for two more budget cycles.

Until the AI layer only works on one side of the gap. And the same documentation says Microsoft is consolidating purchase options and retiring the Power BI Premium per-capacity SKUs, and that new and existing customers should consider Fabric capacity instead.

So the licensing herds you to Fabric, and the AI features herd you to Fabric.

Two ropes, same direction. Though to be fair, there's a real engineering argument underneath. If the semantic layer answers natural language questions, it has to be consistent. You cannot ground a model on forty contradictory definitions.

Sure. And that argument also happens to justify moving every customer onto one stack. Both can be true.

Both are true. That's the honest read. Confirmed is the tooling. A judgment call is how fast you move.

So what do I actually do Monday?

Inventory. Every semantic model with more than, say, fifty users. Two piles. Endorsed... this model may speak to Copilot. Quarantined... it may not, yet. Then use the new controls to enforce the split. You can do that today.

And the fragmented Tabular models out in the divisions?

Don't rebuild them. Pick the three measures that show up in every board pack, define them once, make everything else reference that. Standardization that starts with a thousand-model rewrite never finishes. Standardization that starts with gross margin usually does.

Watching next week?

Two things. Whether Microsoft ever attaches a date to Azure Analysis Services, because that word currently is doing enormous work. And how the remote authoring MCP server behaves in preview, because that's the door between your model metadata and every AI agent in the building.

And nobody knows yet who's walking through it.

For years, a bad semantic model was just a bad report. Now it's a bad answer, said with a straight face, to someone who will never know where it came from.