Writing Prompts for Copilot Report Pages in Power BI - What Actually Gets a Usable Result

There is a quiet shift happening in how reports get built in Power BI, and most people have not clocked it yet. You can now describe the report page you want in plain English and Copilot will draft it for you. Not answer a question about your data, which is a different feature, but actually lay out a page with visuals, a title, and a sensible arrangement, based on a written prompt. The Microsoft documentation covers the mechanics. What I want to talk about is the part that decides whether you get something useful or something you delete after ten seconds: the prompt itself.

I have watched enough people try this to see the pattern. The first prompt is always something like "build me a sales report". Copilot dutifully produces a page. It is fine. It is also generic, and about sixty per cent of what it put on the page is not what the person actually wanted. They shrug, decide the feature is overhyped, and go back to building by hand. The feature is not the problem. The instruction was vague, so the output was vague. Same as it is with a junior analyst you have not briefed properly.

Why the prompt carries so much weight

When you ask Copilot to generate a page, you are handing it two jobs at once. It has to decide what the page is about, and it has to decide how to show it. If your prompt only answers the first job, Copilot guesses at the second, and its guesses are safe and bland by design. It will reach for a clustered column chart and a card and call it a day, because that is the statistically sensible thing to produce when nobody told it otherwise.

The trick is to do more of the deciding yourself and leave Copilot the assembly work, which is the part it is genuinely good at. Think about the difference between these two:

"Create a sales page"

versus

"Create a page showing total revenue and units sold as cards at the top, monthly revenue as a line chart for the last twelve months, and revenue by product category as a bar chart, with a slicer for region"

The second one is barely more effort to type, and it produces something you can actually use. You have told Copilot the metrics, the visual types, the time frame, and the interaction. It is no longer guessing. It is arranging. That is the whole game.

What to put in a good prompt

Through a fair bit of trial and error with Australian teams getting started on this, a rough recipe has emerged for prompts that land. Name the specific measures and fields you want, using the names as they appear in your model. Say what time period matters, because "last quarter" and "trailing twelve months" produce very different pages. Suggest the visual type when you have a preference, because if you do not, Copilot picks for you and you may not like the pick. And mention the slicers or filters the audience will want, because a page without the right filter is a page people abandon.

You do not have to get all of that into one sentence. Copilot handles a paragraph fine, and honestly a slightly longer, specific prompt beats a short vague one every time. The instinct to keep it terse works against you here.

One thing worth calling out: this leans hard on your field and measure names being sensible. If your prompt says "show revenue by region" and your model has a measure called "Revenue" and a field called "Region", Copilot connects them instantly. If your measure is called "Sales_Amt_FY_v3" and your region field is buried in a table called "DimGeo2", Copilot has to guess what you meant, and it guesses wrong more often. This is the same underlying truth that governs every Copilot feature in the platform. The model underneath decides the ceiling. We end up doing a lot of Power BI consulting work that is really just getting models into a state where these AI features can do their job, and prompt-driven page building is another reason that groundwork pays off.

Where it genuinely saves time

I want to be fair about what this is good for, because it is real. The biggest win is the blank-page problem. Starting a report from nothing is the slowest part of the job, and Copilot getting you to a reasonable draft page in one prompt skips the tedious initial layout. You are then editing rather than building, which is faster and more pleasant.

It is also good for the routine, well-understood report. A standard monthly sales page, a basic operations dashboard, the kind of thing that follows a familiar shape. Copilot has effectively seen thousands of these and produces a competent version quickly. For a consultant or an analyst churning through a backlog of similar requests, that is genuine time back.

And it is a decent teaching tool. Someone new to Power BI can describe what they want and see how Copilot chooses to build it, which is a faster way to learn sensible layout than staring at a blank canvas. We fold this into a lot of our Power BI and AI training now, because watching the tool make choices is a good way to develop your own judgement about what makes a page readable.

Where it still falls short

Now the honest part. Copilot is not a report designer. It produces competent, safe pages, and competent-and-safe is not the same as good. The pages it generates tend to be a bit flat. Everything the same size, no clear visual hierarchy, no thought about what the eye should land on first. A skilled report author makes deliberate choices about emphasis and flow, and Copilot does not really do that. It arranges, it does not design.

It also has no idea about your context. It does not know that in your business the margin number matters more than the revenue number, or that the board only ever looks at three specific metrics, or that a particular chart type has burned people before because it gets misread. All of that lives in your head and has to be steered in through the prompt or fixed in the edit afterwards.

And it will confidently produce a page that is subtly wrong if your prompt is ambiguous. Ask for "profit" when you have three different profit measures and it picks one, silently. So you still have to check the output the way you would check anything, because a wrong number in a nice-looking chart is more dangerous than an obvious mess.

The bland output problem is the one to watch most. Because the pages look finished, there is a temptation to ship them as-is. Resist that. Treat the Copilot output as a first draft that gets you eighty per cent of the way, then spend the time you saved on the twenty per cent that makes a report actually clear: the hierarchy, the emphasis, the titles that say something useful, the removal of the visual it added that nobody needs.

How I would use it

My honest advice is to use it as an accelerator, not an author. Write a specific prompt, get the draft, then edit hard. The workflow that works is: describe the page properly, let Copilot lay it out, then go through and fix the hierarchy, tighten the visuals, correct anything it guessed wrong, and cut whatever is noise. That is a genuinely faster path to a good page than building from scratch, and it plays to the tool's strength while covering its weakness.

For teams standardising their reporting, there is a nice side effect. If you write good prompt templates for your common report types, you get more consistent pages across the organisation, because everyone is starting from the same well-specified instruction rather than each person's individual layout habits. That consistency is worth something on its own.

What I would not do is hand this to people who do not know what a good report looks like and expect the tool to supply the judgement. It supplies the labour, not the judgement. Give it to someone who knows what they want and it is a real speed-up. Give it to someone hoping it will tell them what they want and you get a pile of generic pages nobody trusts.

Prompt-driven page building is one of the more genuinely useful bits of Copilot in Power BI, precisely because it plays to what these tools are good at, which is fast, competent assembly from a clear brief. The value is proportional to how good your brief is. Vague in, vague out. Specific in, and you have skipped the boring part of report building entirely.

If you want help getting your Power BI models into shape so these Copilot features actually work, or you want your team trained to use them well rather than just switched on and hoping, that is squarely what we do. Take a look at our business AI services or get in touch and we will sort it out with you.