We have used Claude every day on Power BI for a few months. Here is what it has changed, where it is unmatched, and what it took to get dashboards most of the way there.
How we work with Claude on Power BI
We use Claude, but most of this applies to any AI coding agent. Over the last few months we have used it extensively on Power BI, and it has saved an incredible amount of time. We described our overall approach in Agentic development for Power BI. This is the practical follow-up: how we work with it day to day.
We work in two ways:
- Claude Desktop connected to Power BI Desktop. This uses Microsoft's Power BI Authoring MCP server, which lets an AI agent create and edit a semantic model that is open in Power BI Desktop. Changes are made in real time. It can burn through tokens, so watch your context window.
- Project files and Claude Code. We save the work as a Power BI project (PBIP), which stores the semantic model and the report as text files, then have Claude Code work on those files directly. If you keep the model open in Power BI Desktop, it works much like Tabular Editor: you accept the changes Claude makes and instantly see any issues they cause.
What it has changed
A lot of our semantic modelling is now done through prompts rather than directly in the model. We describe what we want, Claude makes the change, and we review it.
It has its limits. When the DAX gets really complex, especially when we are manipulating a lot of contexts, we still do that by hand and establish the pattern ourselves.
The single biggest timesaver is repetition. Once the pattern is established, Claude can replicate it, so every repeat of it is far quicker.
It is also great for tracing through someone else's work, from a measure back to its source, through every transformation and calculation along the way.
Where it is unmatched: validation
The most powerful thing we have found is validation, particularly when migrating or making changes. Before a change, we have Claude baseline what the model returns. After it, we run the same checks again and compare.
It does this really well. It is fast and it is thorough, and by fast we mean hours where a human would be spending days.
We have extended this to the dashboards and reports. After a model change, we validate them side by side, comparing what they show before and after. This is the final regression step, and we have started using Claude in Chrome to do it. It is Anthropic's browser extension, and it can click through a report in the browser and read what it shows.
Whatever we validate, we are careful about who decides what to check. We build the validation plan with Claude, rather than letting it come up with its own. If the same agent writes the checks and then runs them, they are not independent of the work they are checking. So we agree up front what is being compared, such as the output of the DAX measures, or the old version against the new one in a change request, and a person reviews the checks before they run.
Documentation that finally gets done
We make sure every measure is described in the model itself when it is built, so documentation is a breeze. It used to take us ages, or we never got around to it. Now Claude drafts the description as it builds each measure. We do the same for columns and tables, which we often describe too, especially when we know the semantic model will be used by AI agents.
Where it needs help: dashboards
We have started to explore Claude building dashboards. Out of the box it does an okay job. As with everything, the result depends on the clarity of the brief. Give it something vague and you will get something vague back. Give it real insight into what you want to achieve and you will get a pretty good result. The formatting, though, was not great.
Reports also have fewer tools behind them than the model. Microsoft's MCP server works on the semantic model only and cannot change report pages, so report work goes through the project files, without the live feedback you get on the model. Microsoft's own AI report authoring skills are still in preview.
What made the difference: a design system, not just a theme
We ran it again after investing in a solid design system for the Power BI report. That goes well beyond a theme. It includes page layouts, menu structures, and example visuals for everything we want to use. With it in place, Claude produced something we would say was around 80% of the way there. That is perfect for rapid prototyping.
Things to bear in mind
- The local server does not support macOS. That is rarely a problem for anyone who lives in Power BI Desktop, which only runs on Windows anyway. Microsoft also offers a hosted version, currently in preview, for models published to the Fabric service.
- Edits ask for confirmation by default. The server asks before applying an edit to a model, and there is a read-only mode. Microsoft also recommends keeping project files in Git so every change can be reviewed and reverted, and backing up before you start.
- Claude in Chrome has its own guidance. Anthropic advises against using it on sensitive workflows.
- It is moving fast. Microsoft's server reached its first major version in late September, and the report format and Microsoft's own agent skills are still settling.
When things became available
- 25 August 2025: Anthropic announced Claude in Chrome as a pilot.
- June 2026: Microsoft previewed AI-powered report authoring skills and the Power BI Desktop Bridge.
- 26 August 2026: Claude in Chrome became generally available on every paid Claude plan.
- 23 September 2026: the server was renamed the Power BI Authoring MCP server, having been known as the Modeling MCP server.
- 25 September 2026: version 1.0 of the server was released.
- September 2026: Microsoft's documentation suggests PBIR, the text-based report format that lets an AI agent read and edit a report, is now generally available and the default report format.
Our verdict: use it, and invest in what you give it
The clarity of the brief and a solid design system decide the result. Give Claude both, and a baseline to validate against, and it will save you a remarkable amount of time.
That is where we are putting our effort over the next couple of months. We are building out Claude's knowledge of how we work:
- Design and build standards. Clear standards for all of our solutions, fully documented in a form Claude can absorb.
- Standard patterns. A library of the patterns we use to solve particular challenges, so a pattern is established once and then repeated.
- The design system. Page layouts, menu structures and example visuals, well beyond a theme.
- Better briefs. Understanding what makes a good brief, and helping the whole team get good at writing one.
Claude is already a core part of our day-to-day work, and we expect it to become an even more powerful part of it.
What others are saying
Early commentary matches much of what we have found:
- Check the DAX. Writers agree that an agent makes assumptions from the patterns already in your model, and that being able to see the DAX it wrote does not prove it is correct. The measure and the business definition still need checking by someone who knows them.
- Keep your work in source control. Changes made through an MCP server can be hard to see and undo, so the usual advice is to save as a Power BI project and commit it, as Microsoft's own documentation also recommends.
- Mind the context window. One published example, using Claude Code, found that the Power BI modelling MCP server alone took up 29% of the context in a new session. It comes from one screenshot, but it fits what we see, and the advice is to switch on only the tools you need.
- Design needs rules. For reports, the common view is that a consistent result depends on giving the agent explicit design rules and keeping a person in charge of the design.
What we have not covered: dashboards in Claude
One thing we have not talked about is dashboards built in Claude itself, rather than in Power BI. Dashboards is a beta template within Claude's Artifacts, where Claude builds an interactive dashboard on top of a connected data source, such as a Power BI semantic model. We have run a small experiment so far, and we will share our broader opinion next month.
Sources
- Power BI Authoring MCP server, Microsoft Learn
- Power BI MCP servers overview, Microsoft Learn
- Power BI Desktop projects and report format (PBIR), Microsoft Learn
- Power BI Agentic: agent skills and tools, Microsoft Learn
- Power BI Authoring MCP server releases, GitHub
- What are artifacts?, Anthropic
- Piloting Claude in Chrome and Claude in Chrome is generally available, Anthropic
