NL

Build your first agentic Power BI report in five steps

🎯 Beginner⏱ Half an afternoon
0%

What you'll learn here

Wondering whether you can actually get an AI agent to help you build a Power BI report? You can, today, on your own laptop, without Fabric Premium. The barrier to entry is lower than most people assume. Below is what you really need, why skills make the difference, and how to take the first real step in half an afternoon.

I’m writing this as a starting guide, not a showcase. The bigger announcements around Fabric Apps and Agent Skills at Microsoft Build 2026 are interesting, but you don’t need to wait: the basics already work, and the people who start now keep the choice over how much they hand off to the AI later.

What you actually need

Three things, all local, all without Fabric Premium.

A model saved as PBIP. You normally save a Power BI model as PBIX, a binary file no agent can read. From now on you save as PBIP: the same model, but unpacked into readable text in the TMDL format. That is exactly what an agent needs to understand your tables, columns and measures, and to make targeted changes through version control.

VS Code as your editor. Microsoft’s free editor, available at code.visualstudio.com. Open your PBIP folder there and you have direct visibility on the TMDL structure. The agent lives in the same editor.

The agent itself. Two common options: GitHub Copilot, built into VS Code and activated with your GitHub account, or Claude Code from Anthropic, which works in the same editor and in a terminal. Both are fine to start with. Pick the one you already have an account for.

That’s it. No Fabric capacity, no XMLA endpoints, no extra licensing. The local Power BI Modeling MCP Server (released by Microsoft in November 2025) is the bridge for the modeling layer, and it runs through Power BI Desktop, not through a Premium stack.

Quick check

Why save your model as PBIP instead of PBIX?

Why skills matter

An agent without a skill works out, from scratch on every prompt, how it should handle your work. You can feel it: one time it picks up a convention well, the next time it doesn’t. The results wobble.

A skill is a reusable instruction pack. You give it once, and that working method comes along automatically every session. Your naming conventions, your way of organising measures, your preference for explicit CALCULATE patterns: all of those move out of your head and into a form the agent can use directly.

That is also where you add value as a BI developer. The expertise stays yours. The application becomes repeatable.

The first skill to grab: PBIR CLI

You don’t have to build the first skill yourself. Kurt Buhler and Maxim Anatsko built the PBIR CLI, a community skill for the report layer. You’ll find it at github.com/maxanatsko/pbir.tools.

Installing it is enough to let your agent work with your report files, without standing up a separate server or bridge. Start there, see what it gives you on your own reports, and build out from that.

If you want to let the agent edit the semantic model later, you add the official Power BI Modeling MCP Server on top. That one acts as a local bridge to your TMDL. None of that has to happen on day one.

The plan

Half an afternoon is enough, in this order:

1

Save your model as PBIP

Open a new or existing model in Power BI Desktop and choose File → Save As → PBIP. Your binary .pbix becomes a folder of readable TMDL text files. Start with a new or small existing model; you don’t need to rebuild everything first.

2

Install VS Code

Microsoft’s free editor is available at code.visualstudio.com.

3

Pick your agent

Activate GitHub Copilot from the VS Code extensions with your GitHub account; install Claude Code via Anthropic.

4

Install the PBIR CLI

This community skill by Kurt Buhler and Maxim Anatsko is available via the link above.

5

Start with a small task

Open your PBIP folder in VS Code and give the agent something whose result you can easily verify. For example: add descriptions to every measure in a table, or rename columns to match your naming convention.

📌
Takeaway

The first verifiable task is more important than the most exciting one. You want to see how the agent handles your work before you point it at anything harder.

What stays with you

The finishing touch stays manual. Complex DAX, visual polish, the choices that sit on top of business logic: that’s what you get paid for, and that’s where your judgement makes the difference. An agent produces plausible code, but plausible is not the same as correct, especially around filter context, CALCULATE patterns and ALL/ALLEXCEPT. The stronger your Power BI fundamentals, the sooner you spot when the agent delivers something that sounds right but lands wrong.

Skills cover off a good part of the risk, but only if the basics are in place. That isn’t a barrier to starting; it’s a nudge to build both at the same time.

A look at Fabric Apps

At Build 2026, Microsoft announced Fabric Apps for Semantic Models: AI-first, generation fully handed over. For some use cases that will almost certainly become a beautiful route.

PBIP with an agent in VS Code is a different way in, one where you stay in the driver’s seat. Whoever lays this foundation now gets to choose later how much they hand to the AI. The nice thing is: you can start today.

Final note

Want to take this on with your team in a structured way? Happy to think it through with you. More on the workshops and sparring tracks at trainerbjorn.nl.

Created by Björn, with support of AI, owned by Dogoda. More disclaimers, here.