Finding Data and Asking Questions with Copilot in Power BI - A Practical Walkthrough

Here's a question I hear in almost every Power BI health check we run: "Where's the report that shows X?"

Not "what does the data say". Just where is it. Which of the 340 workspaces, which of the four reports with "Sales" in the name, which one is the current version. In a lot of Australian organisations, finding the right report is a bigger daily obstacle than reading it. People email the BI team, the BI team sends a link, and the same question comes in again three weeks later from somebody else.

The part of Microsoft's Copilot tutorial series covering discovering data and asking questions goes after exactly this problem, then moves on to the question-answering part most people associate with Copilot. Having watched a fair few teams try both, I think the discovery side is underrated and the question side is overrated, at least at first.

Where this lives

This is the standalone Copilot experience. Open the Power BI service at app.powerbi.com and select Copilot in the left navigation. It isn't tied to a single report the way the Copilot pane inside a report is. It can reach across everything you have access to.

That "you have access to" part is important. Microsoft is clear that interactions are specific to each user and only include content the user has permission to see. Copilot doesn't widen anyone's access. But it does make whatever access they already have far easier to use. If your workspace permissions have grown messy over the years, and most have, Copilot will help people find reports they technically could always open but never knew existed. Some of those reports will be ones you'd rather they didn't find. Review who can see what before you promote this experience widely.

Discovering data

The tutorial starts with a fuzzy request: "Where can I find data about the sales funnel including opportunities, stages, and goals". You don't need the report name or the workspace.

Copilot matches on a mix of metadata. According to the docs that includes item names, descriptions, visual titles, text boxes and workspace names, plus signals like your recent views, endorsements, favourites and how popular an item is across your tenant. It searches semantic models, reports, and reports inside apps.

Read that list again, because it's a to-do list for your BI team. Copilot can only match a report to a vague question if the report describes itself. A report called "SalesRpt_v3_FINAL" with no description, untitled visuals and no text boxes gives search almost nothing to go on. A report called "Sales Pipeline - Opportunities by Stage", with a one-line description and proper visual titles, gets found.

Endorsement matters too. Promoted and certified content gets a boost, which means the effort you put into an endorsement process now affects what people see first. If you have no endorsement process, Copilot ranks on popularity and recency, and in my experience the most popular report is not always the correct one. It's often the old one everyone bookmarked before the replacement shipped.

I wrote more about this in a post on searching for content with Copilot. The short version: a weekend spent writing descriptions and endorsing your key reports will do more for Copilot discovery than anything you could configure.

Summarising a report

Once Copilot returns a list, you can ask it to summarise one. The tutorial's prompt is "Provide a detailed summary of the contents of the second item. I need to understand what it contains so I can ask questions."

(Small thing: the docs warn the result order can change between sessions, so "the second item" might not be the one you meant. Use the report name when it matters.)

The summary covers multiple pages and visuals and includes footnotes pointing to the visual each insight came from. From there you can open the report, or use Explore answer to work with the data further and save the result as a new item in a workspace.

I like the summaries more than I expected to. For a manager who has been sent a link to a twelve-page report and has ten minutes before a meeting, a footnoted summary is a real help. The footnotes are what make it trustworthy. You can click through and see the actual visual, so you aren't taking Copilot's word for it.

One practical tip: select Clear chat when you move on to a different report. The tutorial does this between steps, and it's a good habit. Leftover context from an earlier report can bend later answers in ways that are hard to spot.

Asking questions

Now the main event. You ask a question, and if the answer isn't already in a visual in the report, Copilot queries the semantic model and returns a new visual. It applies filters for relative dates and other conditions in the question, and can generate DAX queries for ad hoc calculations that the model doesn't have measures for.

The tutorial's flow has a step I'd tell everyone to make a habit: before asking, select Add items, choose Reports, and attach the report you want from the OneLake catalog. That grounds the question in a specific source. Without it, Copilot has to decide which model to use, and if you have several overlapping models, that's another place for things to go sideways.

Then the question: "What was the revenue for each state in the last year?" You get a short written summary and a visual.

The most useful control on the screen is the "How Copilot arrived at this" expander. It shows the data used and the filters applied. In the tutorial's example, a relative date filter was applied for "in the last year".

For Australian businesses, that's exactly where I'd look first. "The last year" has at least three plausible meanings here: the last twelve months, the last calendar year, or the last financial year ending 30 June. Copilot picks one. Unless the model owner has written AI instructions about your financial year, it won't necessarily pick the one your CFO means. The answer looks equally confident either way. Expand the panel, check the filter, and if it's wrong, rephrase with explicit dates ("FY25, 1 July 2024 to 30 June 2025").

The non-determinism problem

Microsoft includes a note that I wish were larger. Copilot outputs are non-deterministic, so two people can ask the same question of the same data and get different outputs. The docs advise setting expectations with users about what to expect and how to validate what they get.

In practice, that means:

Copilot is great for "roughly how are we tracking", for finding where to look, and for questions you'll check against a report afterwards. It's not the source for a figure going into a board paper, a regulatory return or a bonus calculation. For those, use the governed report, or have the model owner set up a verified answer so the common question always returns the same approved visual.

When we train teams, we spend real time on this. Not because Copilot is unreliable most of the time (on a well-prepared model it does well) but because when it's wrong, it doesn't look wrong. The skill users need is the habit of checking filters and cross-referencing anything that matters, more than clever prompting.

What works and what doesn't

My honest scorecard after watching a number of rollouts:

Discovery works well when reports have decent names and descriptions. It's one of the quickest wins Copilot offers in Power BI, and it cuts the "where's that report" traffic to your BI team.

Summaries work well and the footnotes make them checkable. Good for busy managers and for people new to a report.

Ad hoc questions are mixed. On a clean model with sensible names, a trimmed AI schema and good instructions, they're impressive. On a model built years ago for one set of visuals, they're a lottery. The difference is nearly always the model, not Copilot.

Generated DAX is handy but needs review for anything you plan to keep. Treat it like a junior analyst's first draft.

Getting your organisation ready

If you want this experience to land well, the work is mostly unglamorous:

1. Clean up permissions so people only find what they should.
2. Give your top fifty reports proper names, descriptions and visual titles.
3. Set up endorsement, and certify the reports that are the source of truth.
4. Prepare the semantic models behind those reports for AI.
5. Train users to attach a report before asking, and to check the "How Copilot arrived at this" panel every time.

None of that is new advice. It's good BI hygiene that Copilot happens to reward. The organisations getting real value from Copilot in Power BI are almost always the ones that were already looking after their content.

If you'd like help working through that list, our Power BI consultants do this regularly, and our Copilot training covers the user side with your own reports and data. For a broader view of where AI fits in your reporting, have a look at our work on AI for business intelligence.

The tutorial step itself is on Microsoft Learn: Copilot in Power BI tutorial - Discover data and ask questions.