Copilot in Power BI Explained (2026): What It Does, What You Need and How to Prepare Your Model
Gulab Chand Tejwani
PMP · PL-300 · Microsoft Certified Trainer
· 12 min read
Copilot in Power BI is Microsoft's generative AI assistant built into Power BI Desktop and the Power BI service. It can answer questions about your data in plain language, summarise reports, create report pages, write DAX queries and suggest measure descriptions. To use it, the report or model must sit on a paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher); a Pro or Premium Per User licence on its own isn't enough. The quality of its answers depends heavily on how well the semantic model is prepared.
I'm Gulab Chand Tejwani, Microsoft Certified Trainer. Every Power BI batch I've taught this year has asked some version of "will Copilot replace the analyst?". After using it on real models, my answer is no, but it does change what a good analyst spends time on. The model matters more than ever. Below is what Copilot does today, what it needs, and how to set a model up so it gives sensible answers. Details are from Microsoft Learn as of November 2026; Copilot changes often, so check the docs for anything you're about to rely on.
What is Copilot in Power BI?
It's a set of AI features, most of them chat-based, that use your semantic model as context. When a question relates to data in the model, Copilot answers from the model. Otherwise it may answer from the language model's general knowledge, which is worth knowing when you judge its replies.
Microsoft groups the features by who uses them.
Business users (report viewers) can ask questions about a report in the Copilot pane, get a summary of a page or the whole report, and add summaries to email subscriptions. There's also a full-screen standalone Copilot, still in preview, that searches across every report and model you can access, and an app-scoped version (also preview) that stays inside one app's content.
Report authors can create and edit report pages from a prompt, add a narrative visual that summarises the page, summarise the underlying model, write DAX queries and generate descriptions for measures.
Semantic model authors get Copilot in web modeling (preview), which flags unclear naming or structure, suggests best-practice changes and can help edit measures, tables and relationships.
What do you need to use Copilot in Power BI?
This is where most people get stuck. Copilot runs on capacity, not on your personal licence. There are two ways to get there. The simple one: the report or model sits in a workspace on a paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher); trial capacities and free SKUs don't count. The other: your admin sets up a Fabric Copilot capacity (itself F2 or P1 or higher) and assigns you to it. Your Copilot usage is then billed there, and you can use Copilot in Power BI Desktop and on content in Pro, PPU or trial workspaces without moving them. Microsoft says plainly that a Pro or Premium Per User licence on its own isn't enough.
A few more conditions apply:
- Your admin must leave the tenant setting for Copilot and other Azure OpenAI features switched on. It's on by default.
- The capacity has to be in a supported region. If your tenant is outside the US or EU data boundary (most Indian tenants are), Copilot stays off until the admin allows data to be processed outside your geographic region.
- In Power BI Desktop you need write access to a workspace on a qualifying capacity, because that's where Copilot usage gets billed, unless you're assigned to a Fabric Copilot capacity.
So if you're a learner on a free account or a Pro trial, you probably won't see Copilot working. That's not a reason to skip it: the PL-300 outline mentions Copilot, and employers increasingly ask about it. Learn what it does and how to prepare models for it, and practise on a work capacity when you have access.
How do you prepare a semantic model for Copilot?
Microsoft says it plainly: without preparation, Copilot can misinterpret the data and return generic or inaccurate results. Power BI has a Prep data for AI button (Desktop home ribbon, and the semantic model page in the service) with three tools:
- AI data schema. Choose which tables and columns Copilot should consider. Hiding helper columns, technical keys and duplicate fields cuts confusion a lot.
- Verified answers. Pick a visual and attach trigger phrases to it. When a user asks one of those questions, Copilot returns your approved visual instead of improvising.
- AI instructions. Plain-language guidance about your business: "Revenue means net sales after returns", "Our financial year runs April to March", "Use the Date table, not OrderDate, for time questions".
When the model is ready, the owner can mark it Approved for Copilot in the model settings in the service. Microsoft notes the Prep data for AI features need Q&A enabled on the model, and that changes can take a little while to show up in Copilot's answers.
The basics matter just as much as the AI-specific tools, and they're the same things that make a model good for humans:
- Clear table and column names (
Sales Amount, notamt_2). - A star schema with a proper date table.
- Measures for every number people ask about, with descriptions.
- Hidden technical columns and keys.
- Synonyms for terms people actually use ("turnover" for revenue, "region" for zone).
| Area | Helps Copilot | Hurts Copilot |
|---|---|---|
| Naming | Sales Amount, Customer City | amt_2, cust_cty, Column1 |
| Structure | Star schema, one date table | One wide flat table, several date columns |
| Calculations | Named measures with descriptions | Implicit sums on raw columns |
| Visibility | Keys and helper columns hidden | Everything visible |
| Business rules | Written as AI instructions | Known only to the analyst |
What can Copilot do well, and where does it struggle?
From my own testing on training models, Copilot is useful for:
- First drafts of report pages that you then fix.
- Summarising a busy report page into a few readable sentences.
- Drafting DAX queries and measure descriptions.
- Answering simple "what was X by Y last month" questions on a clean model.
It struggles when the model is messy, when a question depends on a business rule nobody wrote down, or when the honest answer is "the data can't tell you that". It can also give different wording for the same question (Microsoft notes AI behaviour is non-deterministic). Treat every number it produces the way you'd treat a new colleague's first report: check it before it goes to a manager.
How do you write good Copilot prompts in Power BI?
Be specific about the measure, the slice and the time. Compare:
- Weak: "Show me sales."
- Better: "Show total sales by region for the last 12 months as a line chart, one line per region."
- Weak: "Why did sales drop?"
- Better: "Compare sales by product category for September and October and list the three categories with the biggest decline."
Microsoft allows up to 10,000 characters per prompt, so you have room for context. In report creation, describing the audience ("a one-page summary for the regional sales head") tends to produce more useful layouts.
Practice questions: Copilot and PL-300 style scenarios
A team has Power BI Pro licences and publishes reports to a workspace on shared capacity. They want to use Copilot in the report pane. What do they need?
- ANothing; Pro includes Copilot
- BPremium Per User licences for every viewer
- CUse a paid Fabric (F2+) or Premium capacity, either by moving the workspace or by assigning the users to a Fabric Copilot capacity, with Copilot enabled by the admin
- DA Microsoft 365 Copilot licence for each user
Show answer and explanation
Correct answer: C
Copilot in Power BI needs organizational capacity: paid Fabric F2+ or Premium P1+, with the tenant setting on and a supported region. That can be the workspace's capacity or a Fabric Copilot capacity the users are assigned to, which also works for Pro workspaces. Microsoft says Pro or PPU alone isn't enough.
Users keep asking Copilot 'What is our monthly revenue?' and get slightly different visuals each time. You want one approved answer for this question. What should you set up?
- AAn AI instruction that says 'be consistent'
- BA verified answer on the approved revenue visual with trigger phrases
- CRow-level security on the revenue table
- DA bookmark for the revenue page
Show answer and explanation
Correct answer: B
Verified answers link trigger phrases to a specific visual, so common questions return the approved result. AI instructions help with business context but don't pin a question to a visual.
Copilot keeps using a technical column called SK_Customer in its answers, which confuses users. What is the best fix?
- ARename the column to Customer
- BDelete the column from the model
- CTell users to rephrase their questions
- DHide the key column and exclude it from the AI data schema
Show answer and explanation
Correct answer: D
Surrogate keys are needed for relationships but shouldn't be visible. Hiding them and excluding them from the AI data schema keeps them working while keeping them out of Copilot's answers. Renaming it to Customer would mislead, and deleting it breaks relationships.
Your company's financial year runs April to March. Copilot answers 'year to date' questions using January to December. What should you do first?
- AAdd an AI instruction explaining the financial year and make sure financial-year measures exist in the model
- BStop users from asking year-to-date questions
- CSwitch off Copilot for the model
- DChange the Date table to start in January
Show answer and explanation
Correct answer: A
Copilot doesn't know your business rules unless you tell it. An AI instruction plus proper financial-year measures (for example TOTALYTD with a 31 March year end) gives it the right context.
Does Copilot change what you should learn in Power BI?
It shifts the weight. Dragging fields onto a canvas is the part Copilot speeds up most. Modelling, clear naming, DAX measures and knowing whether an answer is right are the parts it relies on you for. If anything, I spend more class time on modelling now than I did two years ago.
If you're building those foundations, the Power BI Mastery course covers modelling, DAX, the service and security across 14 modules and 83 lessons, with labs, 966 quiz questions and 8 PL-300 practice exams, for ₹1,999 personal access. Copilot itself needs a qualifying capacity, so the course teaches the model preparation that makes Copilot useful when you have access at work. For the DAX side, start with DAX functions in Power BI with examples, and for the exam, the PL-300 exam guide.
FAQ
Is Copilot in Power BI free?
No. It needs a paid Fabric capacity (F2 or higher) or Power BI Premium capacity (P1 or higher), either the workspace's own or a Fabric Copilot capacity your admin assigns you to. Copilot usage consumes capacity units on that capacity. Trial capacities and free SKUs aren't supported.
Can I use Copilot with a Power BI Pro licence?
Not with Pro alone. Microsoft states that a Pro or Premium Per User licence on its own isn't sufficient, because Copilot requires organizational capacity. Pro users can use Copilot on content in a workspace backed by a qualifying capacity, or anywhere (including Pro, PPU and trial workspaces) once an admin assigns them to a Fabric Copilot capacity.
What is Prep data for AI in Power BI?
A set of tools for model owners: an AI data schema to choose which fields Copilot considers, verified answers that tie common questions to approved visuals, and AI instructions that describe business rules in plain language.
Can Copilot write DAX?
Yes. Copilot can help write DAX queries in DAX query view and suggest descriptions for measures. Always test the result, because generated DAX can be syntactically fine and still logically wrong.
Is Copilot on the PL-300 exam?
The current PL-300 skills outline mentions Copilot in several places, such as using it in reports and to summarise a semantic model. Check Microsoft's study guide for the latest list.