Co-author and co-learner: Claude (Anthropic).
Reading time about 7 minutes. No setup needed for this part.
On a Saturday afternoon I didn’t know where to type the commands from the Fabric Apps tutorial. By Sunday evening I had two working apps in Fabric: a supplier portal on a lakehouse, and a sales forecast app on a Power BI semantic model, with shared scenarios, an approval step and charts in IBCS notation.
I’m a Power BI consultant, not a developer who lives and breath coding 😉. This series is for everyone who comes from the BI side and wonders: does this concern me, and how do I get started? Short answer to the first question: yes. The second one takes nine posts, and I’ll show you every stumbling block along the way.
The whole series follows one concept: Human + AI co-programming. I don’t write the code myself. I bring the business knowledge and the decisions, and an AI writes the code with me. More on how that works further down.
Where Power BI stops
Power BI is brilliant at showing data. As soon as users want to give something back, it gets painful. Yes, even with the new UDF in Power BI.
You know the situations from projects: sales should enter their forecast next to the actuals. The controller maintains an account mapping that someone copies into the model every month. A department wants to comment on variances or approve an order. In the end, it lands in an Excel file that travels by email, or in a Power App that becomes its own project with its own logic.
Translytical write-back with User Data Functions already made this gap smaller. Fabric Apps go one step further: real applications with their own interface, their own database and sign-in, directly in Fabric, built on top of the semantic model you already have.
What Fabric Apps are, in BI language
A Fabric App is a new item in your workspace, like a lakehouse or a semantic model. Inside sits a small web application built on Microsoft’s Rayfin SDK. What makes it interesting for us is what Fabric provides out of the box:
- Your semantic model as the source. The app queries your existing measures with DAX. No second logic layer, no copied numbers: in my build, Sales YTD in the app matched the source data to the cent.
- Its own SQL database for everything users enter. You describe the tables in code, Fabric creates them and the API that reads and writes them.
- Sign-in with Microsoft Entra ID. No separate login, no passwords.
- Hosting. The app runs at its own address in Fabric; nobody has to manage a server.
The code is TypeScript. That sounds scary, but in practice an AI assistant like GitHub Copilot writes it. Our role shifts to where we are already strong: understanding the data model, describing requirements precisely, and checking whether the numbers are right.
What I built in one weekend
The main app of this series is a sales forecast tool on top of a plain Power BI semantic model: three CSV files, a date table, a handful of measures.
- Actuals and a forecast suggestion come live from the model’s measures.
- Planners overwrite the suggestion per region, product group and month; the app saves it in its own database.
- Scenarios work the way finance teams work: plan a Base, set it to Final, copy it into a new version, re-plan, compare.
- All charts follow IBCS notation: actuals solid, previous year light grey, forecasts hatched, variances in green and red.
From an empty folder to the app above took about three and a half hours of building, plus a preview bug I reported to Microsoft along the way. A version with a lakehouse instead of a semantic model took about three hours the day before; it becomes the bonus part of this series.
Short preview of the working Fabric App
Use cases almost every customer has
The best fits combine three things: read existing data, store user input, and serve internal users who already sign in with Microsoft.
- Planning next to actuals: budget or forecast entry right beside the numbers from the semantic model, like the app in this series.
- Master data maintenance: mapping tables, product groups, account assignments as a form with permissions instead of an Excel file.
- Comments and explanations: explaining variances where they show up.
- Small workflows: requests, approvals, action trackers with status and overview. My supplier portal is one of these.
Read, select, save, output. That pattern carries over to a surprising number of customer questions.
Why now: what FabCon Europe changed
Fabric Apps have been in public preview since June 2026. That alone would make them a nice side topic. What made me start this series is what Microsoft announced at FabCon Europe in Barcelona:
- Pro and Premium Per User support is announced for the coming weeks. Right now you need a Fabric capacity (a trial works). Once Pro and PPU are supported, Fabric Apps stop being a „big Fabric customers only“ topic and reach the many companies that still run plain Power BI.
- Agentic app creation in Power BI Desktop (preview, also announced for the coming weeks). Building an app may move into the tool we already live in.
- Microsoft openly positions Fabric Apps as the next generation of report development. Whether that holds or not, our customers will ask us about it.
⚠️ One honest warning: everything here is preview. Commands, menus and limits will change. Every post in this series carries the date and the CLI version I used, and I will update the posts when things move.
How this series works: Human + AI co-programming
Don’t be suspicious about AI. I got the idea during the weekend because AI was such a great help here and it takes the series I already wrote „AI & Me“ to a next whole level.
This series is built on one idea: Human + AI co-programming. I did not learn to code first and then build the apps. I built them together with AI, and that is exactly what I want to show you.
The roles are simple:
- The human (you) owns the business side: the use case, the semantic model, the measures, the decisions („Base stays approved, use a copy instead“), the testing in Fabric and the final responsibility.
- The AI writes and changes the code, runs the tests and explains what it did. In my case that was GitHub Copilot in VS Code for the app code, and Claude as co-learner and co-author for planning, troubleshooting and writing these posts.
That is why every tutorial step has two blocks:
- You do: what you click, type or check yourself, with an estimated time.
- Prompt for your AI: the prompt I actually used, so your AI assistant gets the right context (model name, tables, measures) instead of guessing.
The posts are written for both readers: for you, and for the AI you paste them into.
Honest assessment: what you need, what it costs, what does not work yet
What you need
- Node.js, VS Code, Git and an AI coding assistant (I used GitHub Copilot, Free plan). No prior coding knowledge, but you need to be fine with a terminal.
- Two tenant settings: „Fabric apps (preview)“ and, for semantic models, „Semantic Model Execute Queries REST API“. Watch out: if a setting is scoped to a security group, being a Fabric admin does not get you around it.
- A Fabric capacity for now (a trial is fine).
What it costs … just your soul and your first born …just kidding its more then that 😈
- Building two demo apps used about 81% of my monthly Copilot Free allowance. What costs credits is the prompts, not the time: a small fix was about 1 point, a small feature about 2, a stubborn layout fix 8, and the whole IBCS styling phase about 19.
What does not work yet (October 2026)
- External users need Entra ID sign-in. There is no anonymous „Publish to web“ for live data.
- The semantic model connector is blocked on the server side for me (Rayfin issue #85). There is a workaround through the Fabric SDK, which I show in Part 4, but it only runs inside Fabric.
- Some rules can only be enforced in the user interface, for example the lock on a final forecast scenario. Fine for a demo, a topic for production.
And the most important point: the AI writes the code, but not the responsibility. You still need to check the numbers, the security and the costs. That is where we BI people are strong!
How VS Code looks like (save this for tomorrow!)
What comes next
- Part 1: Why Fabric Apps matter for BI people (this post)
- Part 2: Setup without fear: Node, VS Code, Git, tenant settings and your first deploy
- Part 3: The semantic model as the foundation: building a model your app (and your AI) can understand
- Part 4: Your first Data App: live numbers from DAX, including the connector workaround
- Part 5: Writeback: a forecast grid that saves to the app database
- Part 6: Scenarios: Base, Final, copy and compare, shared for the whole team
- Part 7: Hands on the chart: your first edits without AI (yes, it’s Vega-Lite, the same language as Deneb)
- Part 8: IBCS styling: actuals, forecasts and variances the way controllers read them
- Part 9: Costs, credits and Git as your safety net
- Part 10: From demo to customer project: the production readiness checklist
Bonus: The lakehouse supplier portal, a second app with a different data source.
What would you like me to cover? Questions, doubts and „this will never work in my company“ comments are very welcome. They make the next posts better.
Written October 2026 with Fabric Apps (preview) and Rayfin CLI 1.36.2. Preview features change, so check the date. Co-authored with Claude as co-author and co-learner, in the spirit of Human + AI co-programming.
Sources