DAX UDFs, Agent Skills, and Copilot: not the same thing

Build 2026 shipped three things under the "AI for Power BI" banner. Read the announcements fast and they blur into one claim: AI can now build your semantic model.

The three capabilities do different things and sit at different maturity levels. You're making different decisions about each.

Start here: DAX UDFs have nothing to do with AI and are ready today.

UDFs let you define reusable logic once and call it from measures, calculated columns, visual calculations, and nested functions. The syntax:

Functions are stored in TMDL: version-controlled, diffable, composable. A UDF can call another UDF. IntelliSense works in Desktop and DAX Query View; SSMS doesn't support it for UDFs. The function is invisible to report users. It's code reuse at the model layer.

The GA announcement skips the decision framework. UDFs eliminate duplicated business rule logic across many measures. Calculation groups handle user-facing transformations that a report author applies. Logic that surfaces in a slicer belongs in a calculation group. The same fiscal calendar adjustment copied into thirty measures belongs in a UDF. Choose the wrong one and you have technical debt that's hard to unwind.¹

One naming collision before you start reading: DAX user-defined functions and Fabric User Data Functions share the abbreviation but are unrelated products. Marco Russo clarified this in the comments of his own announcement post. Check which one the article is about before you read it.²

Agent Skills for Power BI lets an agent build a complete semantic model and report from a natural language description or a screenshot. You describe what you need; the agent calls the modeling MCP server and report authoring skill; the output lands in PBIP format, TMDL files you can read, diff, and review in Git before pushing to the Service.

An agent's output as inspectable text beats a binary blob. You can read it, flag problems in a PR, and understand what was built.

The format doesn't tell you whether what the agent built is correct. A TMDL file from an agent is still a TMDL file. If the relationships are wrong, or a measure ignores filter context and returns plausible-but-wrong numbers, the format doesn't catch it. You do. Auditable means reviewable. It doesn't mean reviewed.

Agent Skills is in preview. The architecture is right. Test the output quality before you commit to it.³

Copilot in web modeling is the oldest capability of the three. Microsoft's Build framing made it look like a new announcement.

Copilot in web modeling has been in preview since before Build. It didn't appear in the dedicated Power BI Build blog. It showed up in one line of a hero post. Copilot can rename tables and columns, adjust relationships, and generate simple DAX in the Service.

It's an authoring assist for structured edits on a model you already understand. The failure mode is the same one it's always had: generating DAX that's syntactically valid but contextually wrong. A measure can ignore an active relationship or calculate the wrong thing in the wrong filter context. No automated check catches it, and the numbers look right until someone notices they're off.

DAX UDFs are GA and ready to ship. Agent Skills has the right architecture: test it in your environment. Copilot web modeling helps with structured edits on a model you understand, and Microsoft previewed it months before Build.

Know which one you're making a decision about.

Next issue: Fabric IQ — what the semantic layer announcement from FabCon actually means, and what it changes for developers building on top of Power BI.

Reusable, parameterized DAX logic, composable and TMDL-stored, now production-ready. Find the business rule you've copied into the most measures and write one function for it. Everything downstream updates when the rule changes.⁴

Shape Map has been in preview since Power BI was new. It's GA. Custom region maps, improved formatting options, no more preview flag. If your organization has held off because of preview status, that blocker is gone.⁵

The delay between data landing in a Lakehouse and being queryable via the SQL analytics endpoint dropped from minutes to seconds, now in public preview. For semantic models built on Direct Lake over Lakehouse data, this fixes the most common stale refresh complaint: data arrives, the model refreshes, the endpoint hasn't caught up yet.⁶

DAX User-Defined Functions (hub page) — Five articles from Marco Russo and Alberto Ferrari covering syntax, model-dependent vs. model-independent functions, parameter types, and the UDF vs. calculation group decision framework. This is the authoritative depth layer for anything in the GA blog. (SQLBI)

AI-Powered Power BI reporting: From design to deployment with agent skills — The official blog for the Power BI report authoring agent skill. The place to start if you're evaluating Agent Skills. (Microsoft)

The one worth your time this week is the one with no AI in it.

If you know a developer who's been waiting on DAX UDFs to go GA — this is the week to send it.