How to build Power BI reports with AI: a step-by-step guide

AI can build Power BI reports now. Building one your CFO reads the same way every month is a harder ask: actuals against plan, variances where they expect them, and the same rules on page 12 as on page 1.

In this summer's issue of the German finance journal REthinking Finance, Dr. Michael Knapp and Prof. Dr. Gösta Jamin of HWG Ludwigshafen showed an AI agent building a multi-page finance report in Power BI. What they ran into is the problem Zebra BI agentic is built for: it's a public agent skill that gives your AI agent the reporting rules and the visuals from the start.

Want to try it now? Skip to the install steps.

What happened when an AI agent met a reporting standard

Knapp and Jamin set out to show what's technically possible when an agent does all of the work. In Claude Code, with an open-source command-line tool and skills they wrote themselves, they built a multi-tab group report on synthetic data, almost entirely through plain-language requests and the code the agent generated, without clicking through Power BI's visual editor.

Then came the reporting standard. Their first attempt applied their own IBCS skill (International Business Communication Standards, the notation many finance teams use so every report reads the same way) to standard Power BI visuals, and the results were unsatisfactory. According to the article, the breakthrough came once the skill configured Zebra BI visuals instead of bending standard ones. They also had the agent screenshot the report after every change and check it against the request before a person saw it, the same kind of check Zebra BI agentic runs.

Their lessons learned are worth reading too: precise prompts pay off immediately, version control is indispensable, and the output isn't production-ready until an experienced controller has checked it. If your German's up to it, read the whole thing.

Knapp, M. / Jamin, G.: Agentic Coding im Reporting: Wie KI-Agents Finance-Dashboards in Power BI automatisiert erstellen. REthinking Finance 4/2026, pp. 47-54.

Build Power BI reports with AI, using Zebra BI visuals

In Knapp and Jamin's build, two things worked together: Zebra BI visuals for the reporting standard, and a skill that told the agent how to use them. By their account, that one-off investment in a skill pays for itself from the second report. They had to write it themselves. You don't.

Zebra BI agentic gives you that skill ready-made, so your agent starts out already knowing the rules. Run it on a Power BI project that already has your data, say what you want to see, and it builds the pages itself: charts, tables, variance columns and IBCS formatting, all in Zebra BI visuals. Like a good analyst, it looks at your data before it decides anything.

What you get

A real report, not a mockup. You get a Power BI project you open in Power BI Desktop, with every number tied to your data model rather than typed in. No clicking through the visual editor to get there.

Variances without writing variance formulas. Actual vs. plan and actual vs. last year, in absolute and percentage terms. Zebra BI visuals calculate the variances themselves, so the agent doesn't need to write DAX (Power BI's formula language) for them. All you need is one measure per scenario.

One reporting standard across every page. IBCS notation is built in, so actuals, plan and last year look the same way on every page, and variances always read the same way. That's what lets your CFO read page six the same way as page one.

A makeover for the reports you already have. Got a report that works but looks like every other Power BI report? Give it to the agent. It keeps your data model and rebuilds the pages with Zebra BI visuals. That's what you see in the video.

It checks its own work. A setup check confirms your machine and model are ready before you start. As it builds, a checker reads your report files and flags faults Power BI won't warn you about, like a sort that would make a table render empty, or a calculated row that would quietly inflate your grand total. Then the agent opens the report in Power BI Desktop, takes a screenshot of each page and checks what it sees.

You decide how hands-on to be. By default, it decides the structure, builds a first page and shows you, then you steer. Prefer to approve each step? Say "work with me" and it proposes before it builds.

Before you start

Always work on a copy. Zebra BI agentic rewrites the report layer of the project you give it. Before you start, copy the project folder, run the agent on the copy, and keep that copy out of synced OneDrive or SharePoint folders: once Power BI Desktop saves the report, your license key is in its files. Your original stays untouched, and you can compare the two side by side.

You'll need five things:

1. Windows with Power BI Desktop. The agent opens your report in Power BI Desktop to check its work, and Power BI Desktop runs only on Windows.
2. A Zebra BI license or the free 14-day trial. You don't install the visuals yourself: the agent adds them to the report. Without a license they run in the free version, which has limits. A license unlocks everything, including comments in Tables+ and Charts+, for you and anyone you share the report with, the same as any Zebra BI report. One license key covers every visual the agent uses, and so does the free 14-day trial. Already have a Zebra BI license on this PC? You're set. If not, Step 2 gets you a trial.
3. Your report saved as a Power BI project (.pbip). The agent works on project files, not on a single .pbix file. Have a .pbix? Open it in Power BI Desktop, choose File > Save as, and pick "Power BI project files (*.pbip)" under Save as type. It's the format Microsoft recommends for development anyway. The setup check in Step 2 tells you if a folder isn't a project.
4. A data model that can show a variance. That means a separate measure for each scenario (for example Actual, Plan and Previous Year, rather than one Scenario column) and a date table marked as the date table. Not there yet? Ask the agent to help write them. The setup check in Step 2 tells you if either one is missing.
5. An AI coding agent that supports agent skills. Claude Code, GitHub Copilot CLI and others such as OpenAI Codex all work, and each needs its own account or plan. The steps below use Claude Code, and the install guide on GitHub covers the others. Never used a terminal? You install the agent once, and after that you type requests in plain English.

Recommended: Node.js, a free download, runs the checker. If your IT team blocks it, skip it: the agent still builds the report and checks the rendered page.

Install it and build your first report

Zebra BI agentic is on GitHub, with the full documentation, the install guide for other agents and a place to report issues. Here's the short version with Claude Code.

Step 1: Install Zebra BI agentic

Open Claude Code and paste these two commands:

/plugin marketplace add zebrabi/zebra-bi-agentic
/plugin install zebra-bi@zebra-bi-collection

Then restart Claude Code. You don't need a GitHub account: the repository is public.

Using GitHub Copilot CLI? Run copilot plugin marketplace add zebrabi/zebra-bi-agentic, then copilot plugin install zebra-bi@zebra-bi-collection. For Codex and other agents, follow the install guide.

Step 2: Check your setup

/zebra-bi:setup-doctor

It checks your machine, your model and your Zebra BI license, and tells you how to fix anything that's missing.

No Zebra BI license yet? The setup check can open the free trial page for you, or fill it in with you if you'd like (you confirm every field and consent box yourself). Your key arrives by email in a couple of minutes. Paste it into the chat, and the agent activates the visuals straight from the report. Your machine remembers the license.

Step 3: Describe the report you want

Use your own words. For example:

Here's my sales model at C:\Reports\Sales.pbip. Build me a monthly P&L page with variance against plan and last year.

Step 4: Review and steer

It builds the first page, opens it in Power BI Desktop to check how it looks, and tells you what it did and what it left out. Check the headline figures, then tell it what to change.

To get updates later, run /plugin update zebra-bi@zebra-bi-collection. The full documentation is in the GitHub repository.

What you still check yourself

Zebra BI agentic is an early release. Two things stay your job:

• Whether your numbers are right. It reads your report and your model, but it can't know what the numbers should be. Always check the headline figures yourself.
• Whether two periods are comparable. If your actuals stop in April and last year runs to December, every variance is off, and nothing in the files says so. Check that your actuals and your comparison cover the same months.

That's where the REthinking Finance article lands too: the agent-driven build works, but not on autopilot, and an experienced person still reviews the result.

Found something that doesn't work? Tell us on GitHub. A report that came out wrong when the agent said it was fine is the feedback we want most.

What's coming next

Zebra AI Labs gives you early access to what we're building behind the scenes, starting with an MCP connector, so more of the AI tools your team already uses can work with Zebra BI. MCP (Model Context Protocol) is the open standard many AI tools use to plug into other software. More on that soon.

For quick answers, there's Zebra AI

Zebra AI is a separate product. It answers your questions in plain language with Zebra BI charts, in the browser or as a Microsoft Fabric workload. When the answer needs to become a report everyone uses every month, use Zebra BI agentic. When you need an answer now, use Zebra AI.

See it live on October 27

Join Dhaval, Luka, and Mark from Zebra BI for a live webinar: AI + Power BI + Fabric: Reports Your CFO Can Actually Read on Tuesday, October 27, 6 pm CET / 1 pm EDT / 10 am PDT. They'll rebuild an existing Power BI report live with Zebra BI agentic, show Zebra AI in Microsoft Fabric, and answer your questions.

Frequently Asked Questions