Copilot in Power BI

What is Copilot in Power BI?

Copilot in Power BI is the assistant that Microsoft has built into Power BI Desktop and the Power BI service. It gives you a chat box next to your data, and depending on where you open it and who you are, it will summarise a report, answer a question about the numbers, write a DAX query, or put together a whole report page for you.

The most useful way to think about it is this. Copilot is an assistant for the work you were already doing in Power BI. It is not a separate product that you buy, it is not a data platform, and it does not reach any further than the things you already have access to. When your question touches the semantic model, Copilot answers from that model. When your question does not touch the model, it answers from the general knowledge of the language model instead. That is a difference you want to explain to your users yourself, before they find out by accident.

Copilot is switched on by default on eligible capacity, and an administrator can turn it off in the Fabric admin portal.

The three places you meet it

The Copilot pane inside a report
This is the panel on the right hand side of a report you have open. It answers questions about that report and the model behind it, and it can summarise a page or a topic for you. This one is generally available, and it is what most people mean when they say "Copilot in Power BI".

The standalone experience
This is a full screen chat that you open from the left hand navigation. Unlike the pane, it is not tied to one report. It finds and answers questions about any report, semantic model or Fabric data agent that you have access to. At the time of writing it is still in preview.

Copilot inside an app
You reach this one from the navigation of an app, and it only looks at the content that sits in that app. It supports verified answers that the app author has prepared, so a common question can return a prepared answer instead of an improvised one. This is also in preview.

What it does for each role

For business users
You can ask a question about the data, summarise a report or a topic, find the right content, and include a Copilot summary when you set up a report subscription. This comes closest to asking a follow-up question to the analyst who built the report for you.

For report authors
You can create and edit report pages from a prompt, summarise the semantic model underneath, add a narrative visual to the canvas, write DAX queries, and generate descriptions for your measures. Some of this sits in the Copilot pane and some of it appears inline, right where you are working.

For model owners
Copilot in web modelling goes through your model and points out unclear structures and inconsistent naming, suggests improvements to the design, and edits the model itself. It can refine measures and manage tables, columns and relationships.

What you need before Copilot shows up

The requirement that catches most organisations out is capacity. A Power BI Pro licence or a Premium Per User licence is not enough on its own. You need a paid Fabric capacity of F2 or higher, or a Power BI Premium capacity of P1 or higher, and trial capacities do not count.

There are three more conditions on top of that. The tenant setting that allows Copilot has to be switched on. Your capacity has to sit in a supported region. And sovereign clouds are not supported, which Microsoft puts down to the availability of GPUs.

If your tenant or your capacity sits outside the United States or outside the EU data boundary, then Copilot is off by default until an administrator allows processing outside your geographic region. For a Belgian company on EU capacity this is usually not a problem, but it is worth checking rather than assuming.

One more practical detail. When you buy capacity or scale it up, it can take up to 24 hours before Copilot notices and becomes available to you.

Why data preparation decides the result

The quality you get out of Copilot is mostly a data modelling question, not a prompting question. Microsoft is quite direct about this. Without preparation, Copilot struggles to interpret your data correctly and gives you generic, inaccurate or even misleading answers.

There are three features that exist specifically to fix that. They are grouped together under "Prep data for AI" and they are saved on the semantic model rather than on the report.

• An AI data schema tells Copilot which tables and fields it should prioritise, so it does not wander off into technical columns that nobody reports on.

• AI instructions are free text where you explain your business vocabulary, along with the rules that are obvious to your team but invisible in the schema.

• Verified answers tie a particular visual to a trigger phrase, so a question people ask often returns a chart that you know is correct.

Once your model is in shape, you can mark it as Approved for Copilot. Reports built on that model inherit the status, and the standalone Copilot stops adding its cautionary treatment to answers that come out of it.

What Copilot costs on your capacity

Copilot usage is metered in capacity units and shows up in the Fabric Capacity Metrics app, where the item history page gives you a 30 day breakdown per experience. The usage runs as a background operation, which spreads the load out instead of letting a burst of questions spike your capacity.

Token consumption is what drives the bill. Anything that Copilot sets off afterwards, such as a DAX query, a refresh or an email subscription, is billed separately through its own normal path. If you want one clean line on the invoice, you can route all your Copilot usage through a dedicated Fabric Copilot capacity.

What to watch out for with Copilot in Power BI

The same prompt can give you a cached answer. If you ask exactly the same question on an unchanged model within a rolling 24 hour window, you get the stored answer back. Clearing the chat does not change that. So if you are testing a change to your model and nothing seems to have improved, reword the question or refresh the model.

Old context leaks into new questions. A long conversation carries its history along with it, which means an unrelated follow-up question can be answered in the wrong frame. Clearing the chat when you switch topics genuinely helps.

Using it in another language is not officially supported. Prompts in Dutch or French sometimes give you a sensible answer, but multilingual use is not supported at the time of writing. Take that into account when you plan a Belgian rollout.

It answers within permissions, not around them. That is a good thing, and it means two people can ask the same question and get a different number back. Explain that to your users before somebody reports it as a bug.

It is non-deterministic by design. The same question does not guarantee you the same wording or the same emphasis twice. For a figure that has to be identical every single time, send people to the report instead of to the chat.