2026-07-12 · Season 3
Power BI Copilot - Simply Explained

Tried Power BI Copilot once, got a wrong answer and gave up? This episode explains what Copilot in Power BI can and cannot do for report authors, analysts and beginners, and shows how to phrase questions, prepare your data model and verify results so you get reliable answers.
In this episode
- What Power BI Copilot actually is
- Copilot's three superpowers
- The big gotcha: confident wrong answers
- How to ask Copilot the right way
- What Copilot cannot do
- What makes Copilot work well
- What it means for beginners
- A practical framework for today
Key takeaways
- Copilot in Power BI drafts DAX measures, builds report pages and answers data questions using your tables, columns, measures and relationships.
- It can return confident but wrong and non-deterministic answers, so use the exact terms from your report and always click through to verify the source visual.
- Invest in clean names, clear relationships and Prep data for AI, and treat every Copilot result as a rough draft rather than a finished product.
Worth quoting
> “Copilot can sound absolutely certain and be completely wrong at the same time. And because it sounds so confident, you might not think to double check it.” > — Mirko Peters (02:59)
> “Don't rephrase into everyday language. Copilot matches vocabulary, not meaning. So a specific question using the report's own language gets you a consistent correct answer every time.” > — Mirko Peters (04:36)
> “Copilot is only as good as your semantic model. Garbage in, garbage out, same as any tool.” > — Mirko Peters (06:23)
> “Users who treat Copilot as a drafting assistant rather than a finished product have much better results. They don't get burned because they're not trusting it blindly.” > — Mirko Peters (07:46)
> “The right approach is to use Copilot as an accelerator, not a crutch. Let it speed you up, but don't let it think for you.” > — Mirko Peters (09:20)
Explore more on M365.FM
- Episode page: the full episode with chapters, quotes and FAQ on Copilot in Power BI
- Power BI – the tool whose built-in Copilot this episode explains, including its limits and pitfalls
- DAX – Copilot drafts DAX measures from plain English and explains each part of the formula
- Semantic model – Copilot reads your tables, columns and relationships, so model quality decides answer quality
- Power BI Desktop – where a Copilot button in the ribbon opens the chat pane for report authors
- Power BI Service – the web version, with a similar chat pane for questions about published reports
- Related episode: Power BI Semantic Models - Simply Explained
Follow and rate M365.FM in your podcast app, and join the free M365 community at m365con.net.
FAQ
Why does Copilot in Power BI give wrong answers?
Copilot in Power BI matches words in your question to words in your report, such as column names, page titles and chart labels, rather than understanding meaning. When your wording differs, it guesses and can confidently pull the wrong figure. It is also non-deterministic, so the same question can produce different answers in different sessions.
How should I phrase questions to Copilot in Power BI?
Use the exact terms already in your report instead of everyday language. If a chart is labeled order volume, ask about order volume, not total orders. Be specific about visual types, fields and time periods, and refine results in follow-up messages, for example asking Copilot to switch to quarterly or sort from highest to lowest.
Can Copilot in Power BI forecast numbers or explain why something happened?
No, Copilot in Power BI cannot forecast future numbers or explain causes. It only works with data already in your model and can show historical patterns, but explaining why something changed needs human business context. It also cannot fix bad source data, and for major formatting changes it gives manual instructions instead of applying them.
Quotes from this episode
Copilot can sound absolutely certain and be completely wrong at the same time. And because it sounds so confident, you might not think to double check it.
Don't rephrase into everyday language. Copilot matches vocabulary, not meaning. So a specific question using the report's own language gets you a consistent correct answer every time.
Copilot is only as good as your semantic model. Garbage in, garbage out, same as any tool.
Users who treat Copilot as a drafting assistant rather than a finished product have much better results. They don't get burned because they're not trusting it blindly.
The right approach is to use Copilot as an accelerator, not a crutch. Let it speed you up, but don't let it think for you.
Apple Podcasts · Spotify · YouTube · YouTube Music · Deezer
Host: Mirko Peters
Guests
Related articles
- Unlocking Business Intelligence: A Deep Dive into Power BI Copilot
- Why DAX Can't Fix a Broken Data Model
- Demystifying DAX: A Beginner's Guide to Data Analysis Expressions
- Mastering DAX Evaluation Context: Why Power BI Isn't Just Excel
- Why Your Power BI RLS Works in Desktop But Leaks in Fabric
- 7 Surprising Facts About Copilot for Power BI
- Mastering Copilot Prompts for Power BI: Practical Strategies and Examples
- Copilot in Power BI Explained
Transcript
Read the full transcript
Today's topic is one almost everyone has heard of but few actually understand, Power BI Copilot. Microsoft markets it as this AI assistant that writes reports for you and on the surface, that's true, but the truth is more useful and a little dangerous. By the end of this episode, you'll know what Copilot actually does, where it falls short and how to use it without getting burned. Here's the thing, most beginners try it once, get a wrong answer, and never come back. You're going to avoid that trap. What exactly is Power BI Copilot? Let's start with the simplest definition. Copilot is Microsoft's AI assistant built right into Power BI. Think of it like having a coworker who knows DAX and report design sitting next to you. You tell them what you need and they help you get it done. Now Copilot uses the same tech as ChatGPT but here's the big difference. It's trained specifically on Power BI features and it's connected to your actual data model. So when you ask it something, it's not guessing based on generic internet knowledge. It's looking at your tables, your columns, your measures and your relationships. You'll find Copilot in two main places. In Power BI Desktop, there's a Copilot button in the ribbon that opens a Chat pane. In Power BI service, the web version, there's a similar Chat pane for asking questions about your published reports. But here's the key thing to know. Copilot can see your tables, columns, measures and relationships. It works with your data, not some generic data set. That's what makes it powerful and it's also what makes it tricky if your data model isn't well-organized. So what can you actually ask this thing to do? Let's break down the three main things it's good at. Copilot's three superpowers. Let's break down Copilot's three superpowers. The first one is probably the most popular. Writing DAX measures from plain English. You type something like create a measure for year-over-year sales growth and Copilot writes the formula for you. It generates the DAX, explains what each part does and gives you the steps to add it to your model. Prompts like, show total sales for last year, produce working time intelligence measures in seconds. Here's the thing. Copilot doesn't add measures directly to your model. It gives you the code and instructions to add it yourself, so you still need to copy and paste. This small price to pay for not having to remember the exact syntax for same period last year. Next up is generating entire report pages from a description. This is where Copilot really shines for report authors. You describe what you want. A sales performance overview with KPIs and a bar chart by region and Copilot builds it. It picks visual types, lays them out on the page, applies your report theme and even adds slices. Users find this is the feature that saves the most time. Instead of spending an hour dragging and dropping visuals, you can describe what you need and get a solid first draft in seconds. And the third superpower is for business users who don't want to dig through reports, answering questions about your data in plain English. Someone can ask what were our top products last quarter and get an answer with supporting visuals. Copilot can also generate narrative summaries. A paragraph that explains what the data on a page actually means. This is perfect for executive briefings or saving time writing report commentary. Instead of staring at a chart and trying to put it into words, you ask Copilot to summarize it and you've got your first draft. The big gotcha, confident, wrong answers. And all sounds amazing and it is. But here's the gotcha people don't talk about enough. Copilot can sound absolutely certain and be completely wrong at the same time. And because it sounds so confident, you might not think to double check it. Let me give you a real example. Someone asked Copilot what were total orders last month. And Copilot confidently returned 14,941. Sounds like a real number, right? The actual number for February was around 2000. So Copilot was off by about 13,000. When the person clicked through to the source, they found Copilot had pulled the all-time total from a headline on a completely different page, not February's figure at all. If you'd put that number in an email or a board presentation, you'd have a real problem. So why does this happen? It's not that Copilot is broken. It's that Copilot doesn't understand your data the way a human does. It's matching words in your question to words it finds in your report. Column names, page titles, chart labels. When you say total orders but your report labels the chart order volume, Copilot has to guess. Sometimes it guess is wrong. There's another issue too. The same question asked twice can give you different answers. Copilot is what they call non-deterministic. The output varies even with identical input. One session you ask, which manager is struggling? And you get a useful breakdown with names and numbers. Next session, same question. And Copilot asks for clarification. What do you mean by struggling? Third session, it gets blocked entirely. Same question, three different outcomes. That's a confusing pattern. For something like dinner ideas, that's fine. But when you're dealing with revenue figures or SLA numbers, you're presenting to your board, you need the same answer every single time. So how do you get consistent reliable answers? The key is in how you phrase your questions. How to ask Copilot the right way? Here's the key insight. Use the exact words that are already in your report. If your report calls something order volume and rag status, then those are the terms you need to type in. Don't rephrase into everyday language. Copilot matches vocabulary, not meaning. So a specific question using the report's own language gets you a consistent correct answer every time. Let me show you what that looks like. A vague question like, what were total orders last month? That's you that wrong answer we saw earlier. But a specific question using the reports own words, what was order volume for March 2026, gets you the correct answer you can actually verify. Same intent, different wording, completely different result. Another tip, be specific about visual types and data fields. Instead of saying show me sales, try create a column chart showing monthly sales for 2024 with data labels. The more context you give, the less room Copilot has to guess what you want. You're basically narrowing the field of possible answers. And here's something a lot of people miss. Copilot supports conversation. If the first result isn't quite right, ask it to adjust. Change the time period to quarterly or sort from highest to lowest. Treat it like a colleague who needs clear instructions. That back and forth is where you get the best results. But there are also things Copilot simply cannot do, no matter how well you ask. What Copilot cannot do. Let's talk about the limits, because knowing what Copilot can't do saves you a lot of frustration. First, it cannot forecast future numbers. If you ask how many orders will we get next month, Copilot will tell you it can't predict the future. It only works with data that already exists in your model. It can show historical patterns as a rough guide, but no predictions. So if you need a forecast, you're still looking at Power BI's native forecasting features or a separate tool. Second, it cannot tell you why something happened. Why did delivery spike in August? Copilot can't answer that. It can show you the numbers, but causal explanations require human business context. Maybe there was a promotion. Maybe a supplier changed. Copilot doesn't know that. It can show you the data, but the story behind it is still your job. Third, it cannot fix bad data. If your source data has errors, missing values or inconsistent naming, Copilot can't magically clean it. Here's the thing. Copilot is only as good as your semantic model. Garbage in, garbage out, same as any tool. If your data is a mess, Copilot will give you confident wrong answers based on that mess. And fourth, it cannot make major formatting or layout changes. It's difficult to change the bar color to blue and it will give you manual instructions instead of doing it. Visual customization still requires you to use Power BI's formatting panels. So for the polished presentation ready stuff, you're still doing the work. So if Copilot has all these limitations, is it still worth using? Absolutely. If you set it up for success. What makes Copilot work well? Your data model matters most. Clean table names, clear columns and proper relationships are what Copilot reads. If your model is messy, Copilot will be messy. There's no way around it. What does that mean for you? Use names people actually say. A field called Revenue is useless to Copilot. Rename it to net revenue. Now Copilot understands when someone asks about revenue. This helps every human who uses the report too. It's one of those fixes that pays off everywhere. There's a feature in Power BI called Prep Data for AI. It lets you simplify the schema and add instructions. You tell Copilot which tables and measures matter. You add context like Geography is used for borrow analysis. That dramatically improves answer quality because you're teaching Copilot how your data works. A small investment of time makes everything Copilot does more reliable. Here's the most practical advice I can give. Start small and validate everything. Don't ask Copilot to build your entire executive dashboard on the first try. Start with one measure or one visual. Verify it. Then expand. Users who treat Copilot as a drafting assistant rather than a finished product have much better results. They don't get burned because they're not trusting it blindly. This is all stuff you can do today. There's a bigger question here about what Copilot means for you as someone learning Power BI. What Copilot means for beginners and citizen developers. So what does all this mean for you as someone learning Power BI? The short answer is that Copilot lowers the barrier to entry. And that's a good thing. You don't need to memorize DAX syntax to start building useful reports anymore. You can describe what you want in plain English and see working examples right away. Think of Copilot as both an assistant and a teacher. It's both at once. It teaches you as you go. When Copilot generates a DAX formula, it also explains what each part does. You see the code, you read the explanation, and over time you start recognizing patterns. You learn what calculate does, how filter works, and how time intelligence functions like same period last year fit together. You're learning by doing which is the most effective way to pick up a new skill. Instead of reading a textbook on DAX, you're writing real measures and seeing how they behave. But here's the part that matters. Copilot doesn't replace the need to understand your data. It can write the formula, but you still need to know if the result is reasonable. It can build the chart, but you need to know if it tells the right story. Validation is still your job. Copilot is a tool, not a replacement for thinking. The risk is over reliance. Beginners who trust Copilot completely are the ones who get burned. That confident wrong answer problem we talked about earlier hits hardest when you don't know enough to spot it. If you've never seen a revenue report before, you might not realize that 14,1441 is way too high for one month. The right approach is to use Copilot as an accelerator, not a crutch. Let it speed you up, but don't let it think for you. Practical framework. How to use Copilot today. So how do you actually use Copilot in Power BI without getting burned? Think of it as a brilliant but forgetful junior analyst. It can draft reports in seconds, but you still need to supervise the work. Here are four principles to keep in mind. First, always verify the numbers. Copilot typically gives you a click through link to the source visual it used. Click that link, confirm the number matches, every single time. If it can't show you a source, be skeptical. That's your first warning sign that something might be wrong. Next, use Copilot for what it does best. Using DAX measures, generating starter report pages, creating narrative summaries, those are its strengths. Complex calculations, causal analysis or anything that requires business judgment, keep those for yourself. That territory is still yours. Also, invest in your data model before you ask Copilot anything. Clean up table and column names, set clear relationships between tables. Use the prep data for AI feature to add instructions. Every hour you spend here pays back 10 in Copilot accuracy. A well structured model is the single biggest factor in getting reliable answers. Finally, treat everything Copilot gives you as a rough draft. It's a starting point, not a finished product. Refine it manually, test it with different filters, and make sure it behaves correctly in all the scenarios you care about. Then promote it to production. That one habit will save you from the most embarrassing mistakes. So here's what you need to take away. Copilot is a powerful drafting assistant for your DAX, reports and data questions. But it comes with a catch. It sounds confident even when it's wrong, and it needs a clean data model and specific prompts to work reliably. Your homework is simple. In Power BI, clean up your main reports column names, and ask Copilot to write one measure you've been avoiding. Most people will just watch this and move on. You have a chance to actually try it and see the difference. If this helped, subscribe and share it with someone starting their Power BI journey.