How To Set Up Power BI Copilot And What It Really Costs

Power BI Copilot is a chat pane that answers questions by querying the semantic model behind your report. Ask it for a chart that isn’t on the page and it builds one. Ask for a calculation nobody wrote a measure for and it writes the DAX itself. Ask it to build a whole report page and it does that too.

I set it up on a real seven-store P&L, 18 months of actuals against budget, and ran it end to end. Here’s what it costs, how to switch it on, and what it does once it’s running.

If you’ve read anything else here on AI for Power BI and reporting, this is the setup piece that should have come first.

Copilot doesn’t come with Power BI Pro

Start here, because this is where most people stop.

Power BI Copilot needs a paid Microsoft Fabric capacity, F2 or higher, or one of the legacy Power BI Premium plans at P1 or higher. A Pro license doesn’t turn it on. Neither does Premium Per User. Neither does a Microsoft 365 Copilot license, which surprises a lot of finance teams who assume the Copilot they already use in Outlook covers this. It doesn’t, and it’s worth reading what your Copilot license includes before you promise anyone a demo. Microsoft is blunt about it in the docs: a Pro or PPU license alone isn’t sufficient, because Copilot requires organizational capacity. Trial SKUs don’t count either.

The good news is the floor came down. This used to require an F64 capacity, which put it out of reach for anyone who wasn’t already running enterprise BI. F2, the entry level capacity, is a couple hundred dollars a month. That’s a number a finance manager can put on a corporate card and test with.

There’s a second reason to care about the timing. Microsoft is retiring the legacy Power BI Q&A feature in December 2026. New Q&A visuals can’t be created, existing ones stop working, and Q&A Setup goes with them. If your reports lean on Q&A, Copilot is the migration path, and there isn’t another one.

One odd wrinkle worth knowing. Copilot runs on the same underlying Q&A engine, so you have to have Q&A switched on for your semantic model before any of the Copilot features work. The front end is being retired and the engine underneath it is being kept.

How Azure, Fabric and Power BI fit together

Three layers, and the names get used interchangeably in a way that confuses everyone the first time through.

Azure is Microsoft’s cloud infrastructure, the same category of thing as Amazon Web Services or Google Cloud. Microsoft Fabric is the analytics platform that sits on top of Azure. Power BI runs inside Fabric as part of that platform. When you buy a Fabric capacity you’re buying compute that Power BI, and Copilot, draw from.

You don’t need to understand any of that to use Power BI. You do need it to get Copilot switched on, because the capacity lives in Azure and the switches live in the Fabric admin portal.

Setting up the Fabric capacity

1. Go to the Azure portal and search for Microsoft Fabric.
2. Create a capacity. Fill in the subscription, the resource group, a name and a region.
3. Pick your size. The pricing runs high at the top end, so read this screen carefully before you click through it. The capacity unit number is the size, and bigger means more expensive.
4. Create it, and give it a few minutes.

One thing that catches people. If you scale a capacity up or buy a new one, Copilot can take up to 24 hours to notice. If everything looks correct and Copilot is still greyed out the next morning, wait a day before you file a support ticket.

The two settings that switch it on

Both live in the Fabric admin portal, and both matter. Miss the second one and the capacity you just paid for sits idle.

The first is a tenant setting called Users can use Copilot and other features powered by Azure OpenAI. It’s on by default, but plenty of admins turned it off in 2024 when Copilot first appeared and never went back. It has to be enabled at the tenant level. Delegating it to the capacity isn’t enough.

The second is the Copilot capacity setting. This is where you point Copilot at the Fabric capacity you just created and say who’s allowed to call it: the whole organization, or specific users and groups.

Read that as a budget decision wearing a permissions dropdown. Copilot consumption is metered against the capacity, so the group you enable here is the group whose prompts you end up paying for.

Once both are set, any Power BI workspace sitting on that capacity is Copilot eligible.

How to tell it worked

Open a workspace on the capacity and look for three things. The Copilot button appears in the report ribbon and isn’t greyed out. A separate Copilot icon appears in the left navigation, which is the workspace-wide version that can reach across every model you have access to. And on the semantic model itself, a Prep data for AI button becomes available.

The same applies in Power BI Desktop once your Desktop file is connected to a workspace on the capacity. Those buttons stop being greyed out and you’re ready.

Asking questions your report was never built to answer

This is the entry level use and the one every reader can run on day one. Open a report in reading view, click Copilot in the ribbon, and the pane opens on the right with a few suggestions. It looks and behaves like the Copilot in the rest of Office. The difference is that it’s querying your semantic model.

Three prompts worth trying, in order, from easiest to hardest:

Summarize this report page in five bullets for a CFO.

That one reads what’s on the page. On my run it pulled the month, revenue, the top contributors by operating income, the revenue against budget trend, the variance themes by location, and the operational drivers. It picked up the last one from the management commentary sitting in the model.

Show me Sales Revenue actuals versus budget by Location for June 2026.

This is where it gets useful, because there was no chart on that page showing actuals against budget by location. Copilot went into the model and built the visual, and it cited every field back to where the data came from.

What was the change in Sales Revenue for Astoria comparing the first six
months of 2026 to the first six months of 2025, in dollars and as a percent?

That third one has no measure behind it at all. Nobody built a first-half-over-first-half comparison in that model. Copilot wrote a DAX query on the fly and came back with an increase of $151,110, or about 16%. That’s the moment the tool stops being a search box.

The filter trap

This one costs people real credibility.

The Copilot pane talks to the semantic model, not to the report page. The slicers you have set on screen don’t pass through. You can be looking at a page filtered to one location, ask for total revenue, and get the company-wide number back with no warning and nothing highlighted in red.

The fix is simple and it has to become a habit. Put the filter in the prompt. Name the location, name the month, name the version. If you want Hell’s Kitchen for June 2026, say Hell’s Kitchen and June 2026 in the words.

Then check the work. Every answer has a How Copilot arrived at this dropdown showing the fields it used, the filters it applied and the exact query it ran. Read it before you repeat a number to anyone. That single habit is what turns a chat answer into something a controller can sign off on.

Building a report page from one paragraph

Switch the report into edit mode and Copilot changes its own suggestions. In reading view it offers analysis and summaries. In edit mode it offers to create a page, suggest content, or answer questions about the data.

There’s a Suggest content for this report button that reads your model and pitches page ideas, which is useful if you’re staring at a blank page. If you know what you want, skip it and be prescriptive:

Build an executive summary page for the CFO's monthly business review. Show
Sales Revenue actuals versus budget by Location, the monthly trend of Sales
Revenue, cost as a percent of Sales Revenue, and cards for total Sales Revenue
actuals and budget. Cover January 2026 through June 2026.

It came back in a few seconds with four charts, the revenue cards, filters across the top, and it pulled the existing report formatting in with it. That’s 20 to 30 minutes of dragging fields, assuming the semantic model is already in good shape. I’ve built an entire Power BI dashboard with Claude before, and the fair comparison is that Claude went further on design while Copilot went faster on anything already sitting in the model.

Then it edits by conversation. I asked for a matrix with location down the rows, months across the columns and revenue variance to budget in the values, which is the view most FP&A teams rebuild by hand every month. It built the matrix but dropped it in the wrong place, so I followed up:

Can you use the new chart to replace revenue and budget by year and month?

It swapped them. Undo and redo both work on Copilot’s actions, which matters more than it sounds, because the first question anyone asks is what happens if it wrecks the report.

Two limits worth knowing before you try this. Copilot handles structure and fields, not styling or themes, so formatting stays your job. And it doesn’t support custom visuals.

Prep data for AI, where the answer quality comes from

Everything above works out of the box. This next part is what gets you from a Copilot that guesses at your business to one that’s been told about it, and almost nobody covers it.

On the semantic model, click Prep data for AI. Three tools live in there, and Microsoft’s own guidance is to work through them before you let anyone loose on the model. Do them in this order.

Simplify the data schema

Everything in your model is selected by default. Deselect the tables and fields Copilot doesn’t need to answer questions. On my model I pulled out the banking tables and the travel and expense data, because I wanted it focused on financial and operational drivers. Two things happen. Every prompt uses less of your Fabric capacity, and answers come back faster because there’s less to read through.

Verified answers

These pin a specific chart to a question. Right-click a visual, choose set up a verified answer, and Copilot suggests trigger phrases based on what the visual shows. I set two on a location scorecard: how much revenue did each location generate, and which locations have the highest operating income. Back in reading view, asking about operating income by location returned that exact chart with a note saying the question matched a verified answer. Same chart, every time, no matter how the question is phrased.

Verified answers save to the semantic model, not to the report, so every report built on that model inherits them.

AI instructions

This is the system prompt for your model. Same idea as instructions in ChatGPT or Claude, written onto the model so every user gets them. Here’s the structure that works, in four blocks:

Business context
[What the company is, the fiscal year, when close finishes.]

Metric definitions
[What variance means. Which direction is favorable on a cost line. What
"plan" and "results" map to in your version column. Any metric your team
says out loud that isn't a field name.]

Analysis rules
[How to slice by default. Which table is authoritative for which question.]

Known data caveats
[The things one person on your team knows and nobody wrote down.]

That last block is the one that earns its keep. My model has a location whose revenue includes internal wholesale sales to the other stores with no elimination entry posted, which means any company total is overstated. I wrote that into the instructions.

Then I asked for total sales revenue across all locations for the first six months of 2026, and added a question on the end: is there anything I should know before I report that number?

It came back with $35,962,652, and then this, unprompted. Before you report it, note that Long Island City revenue includes wholesale sales to other stores and doesn’t record elimination, so this is overstated.

Nothing in the data says that. It came out of the instructions. That’s a control the model owner installs once, and every person who asks a question from then on inherits it. Same trick behind the automated Power BI commentary I run on this model, except the rules live in the model instead of in a script.

The instructions travel too. Ask through the workspace-level Copilot rather than inside the report and it still pulls that context in when it uses the model.

What Power BI Copilot does badly

A review that only lists wins is an advertisement, so here’s the other side.

It doesn’t generate insight. Copilot answers questions about what the numbers are. Forecasting, anomaly detection and key influencers sit outside what it supports. Ask it why revenue dropped in a particular month and you either get a hedge or you get something the language model reasoned out from general knowledge, which is worse.

AI instructions are guidance. A language model interprets them, and there’s no guarantee it follows them every time. Test them, and re-test after you change them.

It ignores your page filters, as covered above. It also caches, so asking the same prompt on an unchanged model inside a rolling 24 hours gives you the same answer back even when you were expecting something different. Clear the chat or change the prompt.

It doesn’t support custom visuals, which rules out a chunk of what people build. If that’s where your reports live, custom Power BI visuals are still a hand-built job.

And it costs money per prompt. Consumption is metered against the capacity, which is exactly why simplifying the data schema is worth doing before you roll this out to a department.

What it costs and who can turn it on

As of September 2026: a paid Fabric capacity at F2 or higher, or legacy Premium P1 or higher. Trial SKUs aren’t supported. F2 showed an Azure estimate of $262.80 a month when I created mine, which is the number on the screenshot above rather than a quote, so price it on the day. A Power BI Pro license, a Premium Per User license and a Microsoft 365 Copilot license get you none of this on their own.

The report Copilot pane is generally available. The workspace-wide standalone Copilot and Copilot inside Power BI apps are still in preview, which means they need their own tenant switch and they aren’t available in every region yet.

If you’re pricing this against the rest of the stack, put it next to the other AI tools for finance your team already pays for, because a lot of what Copilot does here overlaps with things you may be buying twice.

Where to start

If you have a Fabric capacity already, spend 20 minutes on Prep data for AI before you tell anyone else Copilot exists. Write the four instruction blocks, set two verified answers on the charts people ask about most, and trim the schema. A prepped model and an unprepped one are the difference between a tool your team trusts and one they try twice. Once the model is prepped, the same capacity runs a Fabric data agent, which lets your team ask the model questions without opening a report.

If you don’t have a capacity, the conversation with your admin is short. F2, a couple hundred a month, and the two settings named above.

And if the budget conversation stalls, there’s a route that needs no Fabric at all. You can build a Power BI report with AI using MCP straight out of Power BI Desktop on a Pro license.

The AI instructions template I used on this model is in the AI Library, along with the rest of the prompts and templates I use for finance work. It’s free.