Writing Prompts for Copilot Narratives in Power BI - Getting a Summary Worth Reading

Most people meet the Copilot narrative visual in Power BI the same way. They drop it on a page, it spits out a paragraph like "Sales increased over the period with the highest value in March", and they nod, think "that's neat", and never use it again. The paragraph is accurate. It is also the sort of thing anyone could read straight off the chart in two seconds. So the feature gets filed under "nice demo, no real use", which is a shame, because the narrative visual can be genuinely useful once you stop letting it write whatever it wants.

The Microsoft documentation walks through the mechanics of the narrative visual and how to prompt it. What I want to get into is the bit that decides whether the output is worth keeping: how you write the prompt, and what you should and should not expect the thing to do. We have set this up for a fair few Australian teams now, and the gap between a useless narrative and a good one is almost entirely in the instruction.

What the narrative visual actually is

Quick clarification first, because people mix these up. The narrative visual is not the Copilot chat pane where you ask questions about your data. It is a visual you place on the report page, and it writes a text summary of the data in that report or a specific set of visuals. It updates as filters change. So when someone slices the report to Queensland, the narrative rewrites itself to describe Queensland. That is the appeal. A written summary that stays in sync with the numbers, aimed at the person who wants the story without reading every chart.

Left to its own devices, it produces a competent, generic recap. Totals, a high point, a low point, maybe a trend. Fine, but forgettable. The value comes when you use a custom prompt to tell it what to talk about and, just as importantly, what to leave out.

Why the default output disappoints

When you do not give it a prompt, Copilot has to decide for itself what matters in your data. It has no idea. It does not know your business, so it falls back to the statistically obvious: biggest number, smallest number, overall direction. That is why the default narrative reads like a caption. It is describing the shape of the data, not the meaning of it.

The meaning lives in your head. You know that a five per cent dip in one region is a crisis while a twenty per cent dip in another is just seasonal. You know the board only cares about three numbers. You know that "units sold went up" is irrelevant if margin went down. None of that is in the data in a way Copilot can infer, so if you want it in the narrative, you have to say so.

Writing a prompt that earns its place

The prompts that work are the ones that read like a brief to a junior analyst. Tell it what to focus on, in what order, and for whom.

A weak prompt is "summarise the sales data". You already know what that gets you. A prompt that lands looks more like this:

"Summarise sales performance for a regional manager. Lead with total revenue versus target and the percentage gap. Then call out the top and bottom performing regions by revenue growth. Mention any region where revenue fell more than ten per cent compared to last quarter. Keep it to three short paragraphs and do not describe individual months."

Look at what that does. It sets the audience, so the tone is right. It sets the priority, so revenue-versus-target leads instead of some incidental high point. It gives a threshold, so the narrative flags real problems rather than every wiggle. And it sets a length and a boundary, so you do not get a rambling month-by-month recital nobody asked for.

That last part matters more than people expect. Telling Copilot what not to say is often the highest-value instruction in the whole prompt. The default narrative loves to pad. "Do not list every category", "do not describe month-to-month changes", "ignore anything under $10,000" - these cuts are what turn a wall of text into something a busy person will actually read.

The things that make or break it

A few patterns from setting these up in the field.

Reference your measures by name, the way they appear in the model. If you want it to talk about margin, and your measure is called "Gross Margin %", say "Gross Margin %". Vague references like "profitability" leave Copilot to pick from whatever it can find, and it does not always pick the one you meant. This is the same discipline that governs every AI feature in the platform. A tidy, well-named semantic model gives these features something solid to work with, and a messy one makes them guess. A lot of our Power BI consulting work is exactly this groundwork, and the narrative visual is one more thing that gets better once the model underneath is in shape.

Set the audience explicitly. "For an executive" and "for an operations analyst" produce genuinely different summaries. The executive version zooms out, the analyst version keeps more detail. If you do not say, you get a middle-of-the-road version that suits neither.

Give it thresholds, not just topics. "Flag underperformance" is vague. "Flag any region more than fifteen per cent below target" is actionable. Numbers in the prompt keep the narrative honest and consistent as the data changes.

And test it against filtered states. The whole point is that it updates when someone slices the report. So slice it. Filter to a small region with thin data and check the narrative still reads sensibly and does not say something daft. This is where you catch the awkward edge cases before your users do.

Where it genuinely helps

I am generally sceptical of features that write words for you, but this one has real uses.

The strongest is accessibility and glanceability. Some people simply read faster than they interpret charts. A well-prompted narrative at the top of a report gives them the headline in plain English while the visuals sit below for anyone who wants to dig in. For reports that get circulated widely, that is a real improvement in how many people actually get the message.

It is also good for consistency. If you write a solid narrative prompt for a standard monthly report, every version of that report tells its story the same way, in the same order, with the same thresholds. You stop depending on whoever writes the commentary that month being on form. For teams distributing the same report across an organisation, that consistency is worth a lot, and it is one of the practical wins we cover when we run Power BI and AI training for teams getting started with Copilot.

And it saves the analyst the ten minutes of typing up "here is what this report says" that used to go in an email. The narrative does the first draft. You tweak and send.

Where it still falls over

Now the honest part, because there is plenty to watch.

It will state things confidently that are technically true but misleading. "Revenue grew" when growth was one per cent and inflation was higher. It has no judgement about whether a number is good, only about whether it went up or down. You have to build the judgement into the prompt with your thresholds, and even then, check the output.

It does not know causation and will occasionally imply it. A narrative that says sales rose "following the new campaign" when the campaign had nothing to do with it is worse than no narrative, because it reads as analysis when it is coincidence. If your prompt invites explanation, be ready to correct wrong explanations.

Thin data trips it up. Filter to a slice with three data points and the narrative can produce a summary that sounds authoritative about basically nothing. Sparse data plus confident prose is a bad combination, and it is exactly the situation where a casual reader is most likely to be misled.

And there is the general risk of any generated text: because it looks finished, people trust it more than they should. A wrong number in a paragraph is more dangerous than a wrong number in a chart, because the paragraph reads like someone thought about it. Nobody thought about it. A model assembled it from the data in front of it.

How I would use it

Treat it as a briefing tool, not an analyst. Write a specific prompt that sets the audience, leads with what matters, gives real thresholds, and cuts the noise. Then check it against a few filtered states before you ship the report. Used that way, it is a genuinely nice addition to a report, especially one that lands in a lot of inboxes.

What I would not do is drop the default visual on a page and assume the words are meaningful. They are describing the chart, not thinking about your business. The thinking is your job, and the prompt is where you hand your thinking over.

The narrative visual is one of those Copilot features that looks like a gimmick and turns out to be useful, provided you do the work of telling it what you actually care about. Vague prompt, vague summary. Specific prompt with real thresholds and a clear audience, and you have a report that explains itself to the people reading it.

If you want help getting your Power BI models into a state where these Copilot features do something useful, or you want your team trained to prompt them properly rather than just switching them on, that is squarely our patch. Have a look at our business AI services or get in touch and we will work through it with you.