Copilot Narrative Visuals in Power BI - Are They Worth Adding to Your Reports
Every finance team I've worked with has someone who writes the commentary. The monthly report comes out, and a senior analyst spends half a day writing three paragraphs explaining why revenue was down in Victoria, why freight costs jumped, and why the margin looks odd this month. Then it gets pasted into an email or a PowerPoint deck that sits next to the Power BI report.
The Copilot narrative visual in Power BI is Microsoft's attempt to put that commentary inside the report itself, generated on the fly. I've been testing it with clients for a while now, and my view is fairly mixed. It's useful in specific situations, it's oversold in others, and there are a few things you need to sort out before you put it in front of executives.
This post is based on Microsoft's documentation for creating a narrative visual of a page with Copilot, plus what we've seen on real projects.
What the narrative visual actually does
It's a visual you add to a report page, like a card or a chart. Instead of showing numbers, it shows a written summary generated by Copilot from the data in the other visuals on the page (or across selected pages of the report).
When you add it, Copilot gives you a few starting options:
- A general summary of the page
- Key insights for leadership
- Highlights of top-performing or underperforming items
- A custom prompt where you describe what you want
You can also choose which visuals it should consider. That's handy, because a page often has a few decorative visuals or buttons that you don't want summarised.
The important part is that it's dynamic. When a report reader changes a slicer or cross-filters, the narrative can be refreshed to reflect the filtered data. So if a state manager filters to Queensland, the summary talks about Queensland. That's a genuine step up from a static paragraph in a deck.
Each summary also includes references back to the visuals the statements came from, so a reader can check where a claim comes from. I like this a lot. More on why below.
Setting it up
The mechanics are simple once the prerequisites are in place:
- In Power BI Desktop or the service (in edit mode), select the Narrative with Copilot visual from the Visualizations pane.
- Size it on the page. Give it more room than you think. A cramped narrative visual with a scrollbar is annoying to read.
- Pick one of the suggested prompts or write your own.
- Choose the scope: the whole page, specific visuals, or selected pages.
- Generate, review, adjust the prompt, and repeat until it reads well.
The prerequisites are where people get stuck. You need Copilot enabled for your tenant by the Fabric administrator, and the report needs to be in a workspace backed by a paid Fabric capacity (or an eligible Premium capacity). Trial capacities don't qualify. If you're on Pro licences only with shared capacity, you can't use it.
For Australian organisations there's one more setting worth a careful look. Depending on your capacity region and current Azure OpenAI availability, Copilot may need data to be processed outside your capacity's geographic region, and there is a separate tenant setting to allow that. Check what Microsoft currently says for Australian regions before you assume data stays onshore. For clients in government, health or financial services, this needs a conversation with the security and privacy people before anyone flips the switch. We've seen projects stall for weeks because nobody raised it until the demo.
Writing prompts that produce useful narratives
The default prompts are fine for a demo. For real reports, custom prompts make a big difference. A few things we've found:
Tell it who the reader is. "Summarise this page for a regional sales manager who wants to know which products are behind target" produces something much more focused than "summarise this page".
Ask for specific structure. "Give three bullet points, each starting with the metric name, followed by one sentence on the biggest change compared to last month" gives consistent output. Without guidance, the length and style wander around between refreshes.
Name your terminology. If your business calls it "GP%" and the measure is called "Gross Margin Pct", the narrative might use either. Mention the term you want in the prompt.
Keep it short. Long narratives don't get read, same as long commentary emails don't get read. We aim for something a reader can take in within 20 seconds.
Good semantic model hygiene matters here too. Copilot works off the model's metadata, so clear measure names, descriptions and sensible table names directly improve what it writes. A model full of measures named "Measure 7" and "Sales (2) Final" will produce worse narratives. That's been true of every Copilot feature in Power BI, and it's the single biggest factor in output quality that you control.
Where it works well
From what we've seen, the narrative visual earns its place in a few scenarios:
Executive summary pages. A page with four or five KPIs and a couple of trend charts, aimed at people who'd prefer to read a paragraph than interpret a chart. The narrative gives them the "so what" without them having to work it out.
Filtered views for many audiences. One report used by 30 branch managers, each filtering to their own branch. Nobody is going to write 30 bespoke commentaries every month. The narrative gives each of them something reasonable.
Accessibility. For readers using screen readers, a text summary of what the charts show is genuinely helpful. This one doesn't get talked about enough.
First-pass commentary. Some finance teams use it as a starting point. The analyst reads the generated narrative, then writes the final commentary with it as a prompt. That cuts the half-day job down to an hour or so. I think this is actually the most realistic use right now.
Where it falls short
Now the honest part.
It describes, it doesn't explain. The narrative can tell you revenue in Victoria dropped 12%. It can't tell you it dropped because the Melbourne warehouse flooded in week two and orders were held. That context lives in people's heads and emails, not in the semantic model. The commentary your senior analyst writes is valuable precisely because it includes the "why", and the narrative visual mostly doesn't get there.
It can state things with too much confidence. We've seen narratives describe a small, noisy movement as a "significant decline". We've seen it compare the wrong periods when the page had mixed time contexts. The references back to source visuals help a reader check, but most executives won't check. They'll read the paragraph and take it as fact. That's the risk.
Consistency between refreshes. Generate twice with the same filters and you can get slightly different wording or emphasis. For a report that gets screenshotted into a board pack, that variability makes people uneasy.
Cost. Copilot consumes capacity units. On a busy report with lots of readers clicking refresh on the narrative, that adds up and competes with your other workloads. Keep an eye on the Capacity Metrics app after you roll it out.
My recommendation
Use it, but be deliberate about where.
Put it on summary pages for broad audiences, write custom prompts, invest in your semantic model's naming and descriptions, and tell readers plainly that it's AI-generated. Don't put it on a page that goes into a board pack unreviewed. And don't pitch it internally as a replacement for the people who understand the business. It replaces some typing. It doesn't replace judgement.
If you're going to roll it out, do a short pilot with one report and one audience. Ask them after two weeks whether they read it and whether they trusted it. Their answers will tell you more than any feature demo.
We help organisations get Copilot features in Power BI working properly, including the semantic model clean-up that makes them useful. Our Power BI consultants can help with the reporting side, and if you're working out where AI fits more broadly in your organisation, our AI strategy work is a good starting point. For teams that want to get their own people confident with Copilot tools, have a look at our Copilot training.
The narrative visual is a good feature with a narrow sweet spot. Treat it that way and it will be useful. Treat it as automated analysis and it will eventually embarrass someone in a meeting.