Last week I did something that would’ve sounded like science fiction a year ago: I opened Power BI Desktop, typed a sentence, and watched an AI build a complete, themed, multi-page report — measures and all — while I drank my coffee.
No dragging fields onto a canvas. No fiddling with formatting panes. I described the outcome, and Claude did the rest using Microsoft’s new agentic skills for Fabric.
The full build is in the video above. This post is the “why it matters” version — plus the one moment that genuinely surprised me.
What actually got built
Starting from a raw retail model (sales, products, stores, customers), Claude produced a 4-page Power BI report:
Overview, Products, Stores, Customers — with a real top navigation menu
DAX measures written from scratch — including catching that the fact table had no revenue column, so it derived sales as quantity × unit price on its own
A custom theme, KPI cards with icons, sorted gradient charts, and a drill-down matrix
A dynamic AI insight line — “Computers leads at 34% of sales · United States is the top market · 58.6% margin” — that recalculates when you change the year filter
All from plain-English prompts. The kind of thing that’s normally a full afternoon of clicking.
The part that sells it: what happens without the skills
Here’s the demo that made the concept click for me.
I asked a plain AI — no skills loaded — to “build a Power BI page.” It confidently produced JSON that looked right… and opened broken in Power BI Desktop. Invalid visual types. Properties in the wrong place. Error boxes.
That’s the trap with AI and specialized formats: it’s confident but wrong.
The agentic skills fix exactly this. They give the model Microsoft’s actual report spec (PBIR), a validator, and — critically — the ability to reload Power BI and screenshot its own work. So it doesn’t just generate and hope. It generates, checks, sees the render, and fixes itself.
How it works (the 60-second architecture)
There are two channels working in parallel:
Model channel (live): Claude talks to the running Power BI Desktop model over XMLA to create measures and test DAX.
Report channel (files): Claude writes the report’s JSON files to disk, validates them, then reloads Desktop to render them.
Wrapping both is the loop that makes it trustworthy:
write → validate → reload → screenshot → fix → repeat
That self-correction loop is the whole difference between a party trick and something you’d actually use.
Where it got genuinely fun: conversation, not configuration
The best moments weren’t the first build — they were the follow-ups. I just talked to it:
“Make it multi-page with a navigation menu.” → It restructured into four pages with a tab bar.
“The year filter looks off.” → It restyled the slicer.
“Polish the cards and add a dynamic insight.” → Bigger numbers, icons, and that live AI sentence.
And the finale: I dropped a theme file into the folder and said “use this for inspiration.” It read the palette and typography and re-skinned every page to match. Feed it your brand — or a screenshot — and it designs to your taste, not a generic template.
It felt less like using software and more like art-directing a designer who never gets tired.
The honest take
This isn’t magic, and I’m not telling you to ship AI-built reports to your CFO tomorrow. A few real caveats:
You still need to review the DAX and the numbers — the AI is fast, not infallible.
It works best when you can see and verify each step (the screenshot loop matters).
It’s a collaborator, not a replacement — you’re still the analyst deciding what’s worth building.
But the direction is unmistakable. The skill of the future isn’t clicking through menus faster — it’s knowing what to ask for and how to judge the result.
The future of Power BI is agentic. And it’s already here in preview.
Watch the full build
In the video I walk through all of it end-to-end: the announcement, the “without skills” failure, setting up the plugin in Claude, the first build, the conversational refinement, and the design re-skin.
If you found this useful, subscribe — I break down AI + data workflows you can actually use, minus the hype.
What would you hand off to an AI agent first — and what would you never let it touch? Reply and tell me.