How to Build Reports with Power BI Copilot

Power BI Copilot generates report pages, DAX queries, and narrative summaries from plain English prompts. Users open the Copilot pane inside a report, type a request like "create a page showing monthly revenue by region compared to last year," and Copilot selects visuals, maps fields, and builds the layout. The result is a starting point, not a finished report. Refinement through follow-up prompts or the standard Power BI interface typically takes 5 to 10 minutes per page.

Requirements Before You Start

Copilot in Power BI is not available on free or Pro licenses alone. You need one of the following:

• A paid Microsoft Fabric capacity at F2 or higher (the F2 SKU starts at roughly $262/month)
• Power BI Premium capacity at P1 or higher

Trial SKUs do not support Copilot. Your Fabric admin must also enable the tenant setting "Users can use Copilot and other features powered by Azure OpenAI" in the Fabric admin portal. If your organization operates outside the United States or France, the admin must additionally enable cross-region data processing, as Azure OpenAI compute may be served from a region outside your tenant's geography.

Sovereign clouds are not supported as of April 2026.

Step 1: Prepare Your Semantic Model

This step is the most commonly skipped and the most consequential. Copilot reads field names, table names, and any descriptions you add to the semantic model to understand your data. A table named "Fact_SalesOrders_v3" with a column called "Amt_Net_USD" will confuse it. A table named "Sales" with a column called "Net Revenue (USD)" with a description that says "total revenue after discounts, in US dollars" will not.

Before using Copilot on a production semantic model, do the following in Power BI Desktop or the Fabric data model editor:

1. Rename tables and columns to plain business language. "Units Sold" beats "QTY_SHIPPED_NET."
2. Add descriptions to at least your top 10 most queried measures. Select a measure, open the Properties pane, and add a description in the Description field.
3. Mark the date table using Table tools > Mark as date table. Copilot uses this to resolve time-based prompts correctly.
4. Hide columns that are internal keys or calculation helpers. Copilot surfaces hidden fields less often, which reduces noise in generated visuals.

According to Microsoft's documentation, organizations that invest in semantic model preparation see significantly better Copilot accuracy, with accuracy on simple analytical tasks estimated at 85 to 90 percent for well-structured models versus 60 to 70 percent for complex or poorly labeled ones.

Step 2: Open the Copilot Pane

In Power BI service, open any report in edit mode. Select the Copilot button in the ribbon on the right side. The Copilot pane appears as a chat interface alongside your report canvas.

In Power BI Desktop, the Copilot pane requires write access to a workspace backed by a Fabric or Premium capacity in the cloud. Copilot does not run locally. When you open the pane, Desktop connects to the cloud capacity to process prompts.

If Copilot does not appear in the ribbon, verify that the tenant setting is enabled and that your workspace is assigned to an eligible capacity. The most common failure is a workspace still on Shared capacity after an org upgrade.

Step 3: Write Effective Prompts

The most effective prompts describe a business question, not a chart type. Telling Copilot "create a bar chart of sales" produces a generic output. Describing the analysis you want produces something useful.

Examples that work well, drawn from Microsoft's official prompt guide:

• "Create a page to analyze the sales amount, revenue, and profit margin of different products, categories, and subcategories over time and across regions."
• "Create a page to identify and compare the characteristics, behaviors, and preferences of different customer segments based on demographic, geographic, and transactional data."
• "Create a page to evaluate the impact and return on investment of various promotions on sales, revenue, and customer acquisition and retention."

For follow-up refinements, treat it like a conversation. After Copilot generates a page, you can add to the same chat thread: "now add a slicer for product category" or "replace the pie chart with a stacked bar chart." Copilot iterates on the existing page rather than rebuilding from scratch.

For narrative summaries, the prompt format is different. Open the Copilot pane while viewing a finished report page and ask: "Summarize the key insights on this page for a non-technical executive." Copilot reads the visuals and generates a text summary that you can embed as a visual on the report itself.

Step 4: Generate a Table of Contents

For multi-page reports, Copilot can generate a table of contents after you have built all pages. Open the Copilot pane and type: "Make a table of contents for this report that contains a brief description of what each page is about." Copilot reads all existing pages and generates a navigation page with descriptions, which is useful for reports distributed as Power BI apps.

Step 5: Review and Refine

Copilot outputs are starting points. After generating a page, review it for the following:

• Correct field mapping. Copilot sometimes selects a similar-sounding field when the intended field has an ambiguous name. Click any visual and check the fields in the Visualizations pane.
• Date logic. If a prompt includes "last year" or "YTD," verify that the date filter is resolving correctly against your marked date table.
• Aggregation. Copilot defaults to SUM for numeric fields. If a field should be averaged or counted distinctly, adjust the aggregation in the visual's field well.

Any changes made through the standard Power BI interface persist. Copilot actions and manual edits are not in conflict, and you can use undo (Ctrl+Z) to reverse individual Copilot actions if needed.

What Copilot Still Cannot Do

As of April 2026, Power BI Copilot does not:

• Write complex multi-step DAX measures reliably. It can generate simple calculated measures but accuracy drops on measures involving CALCULATE, context transitions, or multiple filter arguments.
• Connect to external data sources on its own. It works only on the semantic model already published to the workspace.
• Apply row-level security dynamically. If your report requires RLS, configure it separately in the semantic model.
• Produce reports from unstructured data like spreadsheets or text files. The data must be loaded, modeled, and published first.

For analysts who work primarily with flat files, spreadsheets, or ad hoc data, the setup cost of loading data into Power BI's semantic layer before Copilot can help is a real constraint. Tools designed around file upload and natural language querying handle that use case more directly. If you want to skip the modeling step entirely, VSLZ AI accepts file uploads directly and runs analysis from a single prompt with no semantic model required.

Practical Summary

Power BI Copilot is most effective for organizations that already have a clean, well-labeled semantic model in Fabric or Power BI Premium. The payoff is real: report pages that previously took a few hours to build can be prototyped in minutes. The gap between good and poor results comes almost entirely from semantic model quality, not from prompt writing skill. Invest 30 to 60 minutes cleaning field names and adding descriptions before your first Copilot session, and most prompts will produce usable output on the first try.