We’re excited to announce the preview of a new chat with your data experience in Power BI. This new standalone, full-screen Copilot experience, easily accessible from the left navigation of Fabric, can help you find and ask questions about any data you have access to. This experience will be available in Power BI Service in the upcoming weeks.
Find the right data when you need it
Your critical business data is at your fingertips in Power BI, but sometimes finding the right report or semantic model can feel like looking for a needle in a haystack. You know you’ve seen it before, you know the general topic, maybe you even remember the name of the visual you’re trying to hunt down. But you cannot seem to remember the name of the report or what workspace you saw it in.
That’s where the new standalone Copilot experience comes in. You can ask Copilot to find reports, semantic models, apps, and data agents that you have access to. We’ll match across many different factors to quickly find the most relevant items.
Metadata, such as the item’s name, description, content (such as, visual titles or textboxes), and the workspace name, are the most important properties used. However, we’ll also consider other attributes such as how recently you viewed it, if it was endorsed, if it’s in your Favorites, and its popularity within your tenant.
We’ll return a hyperlinked list of relevant items with all the details you need to pick which is the one you were looking for.
Key insights only a question away
The standalone Copilot experience doesn’t just find content for you though. It can also help you answer new ad hoc questions and find new insights.
Using Copilot’s summary capabilities allows you to quickly identify the most interesting data within your report. Often, reports can become quite complex, and you can easily spend 30 minutes to a couple of hours combing through all the details. However, Copilot can help you sift through the report, giving you easy to digest overviews of your data. Just ask Copilot to ‘Tell me about trends in sales’ or more specific topics, like ‘What should I know about bike sales in Washington?’.
Copilot will respond with a detailed summary of your report and nest the most relevant visuals right in the reply. These visuals link back to their place in the report but can also be explored so you can dive deeper – applying your own filters or changing the layout.
You may also have a business question that cannot be answered by the existing report. Typically, what you’d have to do in these cases is work with an analyst to get a new visual added to the report, which can take time, delaying your ability to make a data driven decision. This can be especially cumbersome if it’s not something you’d want to track long term.
With the standalone Copilot experience, you can easily ask questions on the semantic model, allowing you to quickly find answers to ad hoc questions and allowing your analysts to focus on critical business requirements.
All you need to do is ask the question and you’ll get a new visual created with the answer. If you know what report or model the right data has, for best results, you can attach it to the chat.
But even if you don’t, Copilot can still help. Just ask your question and Copilot will first find the most likely item you’d want to reference, ask you to confirm which one you’d like to use, and then return the result. You can also ask follow-up questions based on what you’ve already asked in the chat session.
Depending on the type of question you ask, if it’s not in the model already, but could be derived from it, Copilot can also generate DAX queries to answer questions that require ad hoc calculations. For example, you could ask what the year-over-year growth for sales was, or the ratio of a particular category of products to all products.
And of course, just as with the Copilot pane, you can explore your visuals further using our Explore experience, allowing you to seamlessly transition between using chat and traditional drag and drop UI to explore your data.
You can also ask questions against your Fabric data agents. Data agents are AI-powered data assistants that can help determine when to use specific data, how to combine it, and what insights matter most. With data agents, you can also fine-tune results based on your organization-specific instructions, examples and guidance so you can answer data questions based on your data stored in Microsoft OneLake. Just attach the data agent along with your question and you’ll get your answer right in-line. Read more about this integration here.
Check out this demo by Tori where she gives a tour of the new standalone Copilot experience:
High quality, understandable answers
Our goal is that month after month, you’ll see improvements in the quality of the answers you get back from Copilot. For example, in March we made a huge improvement to the language understanding layer, enabling Copilot to understand more of your questions out of the box. In particular, Copilot got much better at understanding common synonyms (ex. tea is a synonym for chai), phrasings (what ‘export’ means in the context of suppliers), and individual values in your model.
We’re taking that a step further this month by improving Copilot’s ability to handle relative dates and to generate date related filters.
| Relative dates by month | Partial dates like ‘March’ uses the current year even if not specified ‘March 2025’ | Relative dates by day of the week |
 |
 |  |
You’ll also notice that in this standalone Copilot experience, the visual answers you’re getting back have a style applied to them. This new styling is meant to make the visual cleaner and help with the readability of the answer.
Simplified data prep for AI
To unlock the full potential of AI, you need to consider the quality of your semantic model. Your data is the foundation upon which consumers will be getting their insights, and the clearer, more structured, and context rich your data is, the better the answers consumers will get. But we also know it can be hard to understand how to influence the quality of Copilot’s answers and test the results. With that in mind, we’re also excited to announce a suite of Copilot tooling features all around preparing your data for AI.
You can easily reach these Prep data for AI features right in the ribbon in Power BI Desktop, and from here, you can easily see the three primary features we’re launching: AI data schema, verified answers and AI instructions.
We recommend starting by simplifying your data schema. Most semantic models out there today are optimized with reporting in mind, and what works best for reporting isn’t always what’s best for AI. For example, you may have measures that are specifically designed to work in the context of a single visual on the report, or you could have many versions of the same measure with different filter context such as ‘sales’, ‘sales in 2015’, and ‘sales last period’. Often, these fields don’t need to be considered for Copilot responses.
Using the Simplify the data schema option, you can easily uncheck fields, hiding them from Copilot, which in turn, improves the relevance, clarity and accuracy of Copilot-generated responses.
After simplifying the AI data schema, we recommend creating verified answers: curated responses that are automatically triggered when users input predefined phrases in Copilot chat experiences. You’ve not only invested time in building up your semantic models, but also building out very detailed reports based on your business needs. Often you already have a visual that answers the most frequently asked questions, but users can’t always find it easily. Verified answers helps solve this. Just mark that visual as an answer to the question, and Copilot will reference it going forward.
Verified answers help improve the quality of Copilot through two ways:
- Improve response accuracy – Surface curated visuals and content to deliver more precise, context-aware answers—especially for nuanced or frequently asked questions.
- Enhance Copilot intelligence – Teach Copilot what a ‘good’ answer looks like by linking specific questions or keywords to trusted responses. Over time, this helps Copilot respond more intelligently to similar prompts.
To get started, just open the ellipsis menu of the visual you want Copilot to use, and pick the Set up a verified answer option.
This launches the setup dialog where you can define trigger phrases, keywords, or full questions that users are likely to ask. Once configured, the associated visual will be returned when a user enters a matching or similar phrase into a Copilot chat. You can also include filters to ensure that users can get filtered versions of the visual in responses.
Additionally, you can manage existing verified answers from the Prep data for AI dialog.
The third feature is adding AI instructions. AI instructions enable you to enhance Copilot’s capabilities by connecting it to your organization’s knowledge. By providing Copilot with key business context, guidance, and domain-specific logic, you can help Copilot understand your business’ specific context and provide more relevant, accurate responses. Whether it’s defining busy seasons, defining how metrics should be interpreted, or excluding irrelevant data, AI Instructions allow you to shape how Copilot understands and interacts with your data, resulting in more tailored insights and greater confidence across Copilot outputs.
All you need to do is go to the Add AI instructions section of the Prep data for AI dialog, add the instructions, and hit apply.
Of course, as you work through using these various tooling features to improve the quality of Copilot for consumers, you will want to test how changes impact the quality of the results. This can be quite an iterative process, and you’ll likely want to jump between testing key questions and the tooling dialog to make adjustments. To give you a little more flexibility while going through the testing process, we’ve introduced a skill picker in the Copilot pane in Power BI Desktop.
This skill picker allows you to focus just on the specific subset of Copilot capabilities you’re working on currently. For example, if you’re trying to reduce the data schema to improve the quality of results when Copilot generates visuals off the semantic model, you may want set the skill picker to just answer questions about the data, so you don’t need to work about Copilot trying to generate a report page based on your questions.
The skill picker includes three capabilities currently, which can be swapped between using the ‘select skills’ option in the area you type your prompt:
- Answer questions about the data
Leverages Copilot to respond to questions based on a given semantic model. - Analyze report visuals
Enables Copilot to interpret and answer questions about the visuals within a report. - Create new report pages
Lets Copilot generate new report pages based on your prompts.
Lastly, once you’ve finished preparing your data for Copilot and testing all the changes, you should mark your semantic model as Prepped for AI. This option is available under the AI Preparation section of the semantic model’s settings in the Power BI service.
We recognize that preparing your data for AI is an important part of the process of adopting Copilot in your organization. This is especially true when using the new standalone, which makes it easier for consumers to explore semantic models and find their own insights.
With that in mind, within the standalone Copilot experience, answers that are coming from models that have not been prepped will be clearly marked. You will get a clear message indicating that the results should be thoroughly evaluated for accuracy and a button that must be selected to view the answer.
Once view answer is selected, the visual will appear inline. To avoid consumers needing to go through this extra friction, make sure to mark your models as prepped for AI once they’ve been tested.
Check out this demo by Anita where she demonstrates how to prep a model for AI:
Data security and compliance
Microsoft Purview data security and compliance capabilities that were announced earlier this year, also extend to the standalone, full-screen Copilot experience. This Microsoft Purview integration helps admins discover sensitive data risks in prompts and responses, detect and investigate risky AI usage (e.g., data exfiltration or policy violations), and govern AI usage through auditing, eDiscovery, retention policies, and non-compliance usage detection.
More coming soon
We’re excited to see how folks take advantage of this new standalone Copilot experience and leverage the continuous quality improvements and the new suite of tooling features. This is the next chapter for Copilot in Power BI, but definitely not the final one. We will continue to improve the experience and are already working on our next set of features.
As a sneak peek of what’s to come, you should expect to see several updates in the next few weeks:
- Refinement on the friction users experience for responses based on unprepped models.
- Textual summaries for generated visuals to help you better understand the main insights of the visual.
- The ability to ask detailed questions about the content of reports, not just summarizations.
- An admin setting to limit Copilot to prepped items by workspace or domain.
Lastly, this standalone Copilot experience is currently in preview, so make sure to turn on the “Users can access a standalone, cross-item Power BI Copilot experience” tenant setting to take advantage of this new feature.
Make sure to try out the new standalone Copilot experience and give us your feedback to help shape our roadmap!