Power BI MCP and its opportunity (to crush its citizens)
The model context protocol (MCP) offers cutting-edge possibilities for Power BI Desktop, but it comes with a few downsides that we will shed light on. As well I'll explain the "clickbaity" title.
Why?
Currently, we see MCP services popping up like snowdrops as soon as the snow melts. As a reminder, MCP is an open source solution, built by Antrophic, that enables LLMs like Claude, ChatGPT or even locally hosted models via Ollama or Llama.cpp to interact with software or services.
In the November 2025 update, Power BI Desktop enables the MCP preview feature where users can use coding editors like Visual Studio Code to interact with their Power BI Reports using natural language.
How?
Through the MCP, the chatbot gets tools like "measure_operations" or "column_operations" to alter the tables and the measures in a way it suits the model (based on the prompt of course). There are a vast suite of tools the chatbot can call via this MCP server, and the possibilities will most likely increase in quantity in the near future.
Each time there is a new requirement to the report, it goes like this:
- The user takes the requirement and "translates" it into a prompt for the connected chatbot
- The chatbot, with all the context it has, tries to understand it and structures the needed steps
- The chatbot calls the tools with created parameters
- The command gets sent to the open Power BI Desktop program and is interpreted by the Analysis-Services-Server that powers the data model and the VertiPaq engine
- The results are immediately visible and can be reviewed
While the chatbot works, the prompter eventually has to intervene and give permission or additional info for the chatbot to continue. So working with Power BI evolves into a trilogue between the developer, the chatbot and the Analysis-Services-Server.
Why this is good:
The MCP enables inexperienced Power BI developers to quickly get into Power BI by seeing what the model does with certain requirements.
Best practices available from the internet has been trained into the LLMs. This means, the reports will somehow be kind of standardized.
The MCP has the potential of enhancing the revision cycles of reports/data models by being able to process bulks of context. Humans may not grasp the full context of measures and tables - LLMs with sufficient context-length can.
The preceding development of this feature likely pushed the development of the PBIR/PBIP file format for Power BI reports. This opens various possibilities to all Power BI developers.
Why this is bad:
Depending on the models, the output and the actions of the model vary vastly with every prompt - every chatbot or LLM writes its code in a different way. It will be harder and harder to understand the data model throughout (for humans).
Users are forced to use Fabric and Microsoft services which leads to higher dependence. Usage may be limited to data that is in the cloud. One way or another, the licenses will get more expensive.
Tokens usage is almost uncontrollable and will lead to higher operational costs, especially if the newest LLMs are used.
Prompt engineering is still not seen as a full profession. The prompts have to be very precise and the user needs to know exactly what he or she wants. This is - unfortunately - not always the case.
Last but not least, models are trained mostly on open data. Enterprise data, that may represent a competitive advantage, is and will not be available on the broad internet. This means, there is and will never be real world training-material for LLMs.
Now to the explanation of the title:
From my point of view, Power BI is merely used by people who on the one hand do deep diving in data but on the other hand still have other tasks in the companies they are working in aka "Citizen Data Scientists". Because of the easiness of creating powerful reports, Power BI was since its creation a tool for practitioners, not for data scientists - they use other tools. Data modeling with M-Code and DAX is the literal core of Power BI. If this part is automated by AI, the bigger part of the Power BI customer base will get eliminated and the focus will shift to the visualization of the data. But to be able to build meaningful reports, one has to understand the data behind it.
If you work with Power BI in such ways as creating underlying data models or building reports, there will come the moment where you will be confronted by the question why your tasks should not be automated by AI. I propose you may prepare yourself to defend against something that will never say 'no' to something your boss tells it.
More info on the Power BI MCP can be found here on Github and this is the blog post where MCP is officially previewed.
More on the "Citizen Data Science" in this post from the Fraunhofer Academy.
Comments ()