Creating Narrative Visuals with Copilot
An executive opens a Power BI sales report and sees a polished paragraph explaining that revenue improved, one region led the increase, and a product category fell behind plan. The paragraph came from a Copilot narrative visual. It sounds plausible, and the report author successfully added it to the page. Is that enough evidence that the solution is correct?
No. Successful configuration only proves that the feature produced some text. It doesn’t prove that the narrative addresses the business requirement, represents the intended visuals, reflects the current filter state, or is suitable for the people who will read it. A useful test starts with three questions.
What is the narrative supposed to summarize? Which Power BI layer owns that behavior? What testable evidence will show that the resulting text agrees with the report?
This matters directly to the PL-300 scope. The course blueprint explicitly includes creating a narrative visual with Copilot. The important skill isn’t simply remembering that the visual exists. You need to understand what it summarizes, how its context affects the result, and why generated text still requires validation.
Suppose a sales and operations team has a report containing revenue, gross margin, units sold, and delivery performance. The executives want a short written explanation beside the charts. They don’t want a new calculated column or another data transformation.
They want readable commentary about results already presented in the report. That requirement points to the report layer, where visuals present and interact with information. A Copilot narrative visual can summarize a whole report, one page, or selected visuals, and a prompt can guide the requested narrative.
That layer distinction prevents a common wrong turn. Power Query performs transformations associated with preparing data for loading and refresh. The semantic model organizes the data used for analysis, including relationships and calculations. Data Analysis Expressions, usually called DAX, can calculate values evaluated under a filter context.
The report layer then presents those values through visuals and interactions. A Copilot narrative visual belongs at that presentation layer. It creates a written summary from available report context. It doesn’t replace missing source data, repair a poor relationship, or redefine an incorrect measure.
Imagine the gross-margin measure is wrong because the underlying model logic is wrong. A well-written narrative can still repeat or interpret that wrong result. The generated prose doesn’t independently certify the calculation. Fixing the narrative’s prompt would also be the wrong response because the cause exists upstream. You would correct the model or calculation, confirm that the visual now displays the expected value, and then refresh the generated narrative against that corrected result.
The opposite distinction matters too. If the measures and charts are correct but the business wants a concise explanation beside them, changing Power Query or adding a calculated field doesn’t directly meet the requirement. The missing behavior is communication at report time.
Selecting the report, page, or relevant visuals as the narrative’s scope connects the written explanation to the intended report content. The nearest technical feature isn’t automatically the right feature. The layer must match the requested outcome.
For the sales report, the analyst first clarifies what the executives need. Perhaps they want an explanation of current-quarter revenue and margin by region, with attention to the most material changes. A report-wide summary might include unrelated delivery and inventory information.
A page summary may be appropriate if the page contains only the intended sales analysis. Summarizing selected visuals gives the analyst a narrower choice when only particular charts should contribute to the narrative. The decision is driven by relevance, not by which option produces the longest paragraph.
Consider what happens if the same page contains a revenue chart, a margin chart, and a delivery-performance chart. A report-level narrative can draw from material beyond that page. A page-level narrative can use all three visuals on the page.
A selected-visual narrative can be limited to revenue and margin. None of those choices is inherently more accurate. They answer different questions because they expose different report evidence to the narrative.
If the monthly review is about commercial performance, including delivery performance may broaden the text beyond the request. If delivery delays are needed to interpret margin, excluding that visual may remove relevant context. The business question determines the boundary, and the resulting prose must be judged within that boundary.
A prompt can guide Copilot toward the requested emphasis, but the prompt doesn’t free the narrative from its available context. If the chosen visuals don’t contain the information needed for the requested explanation, confident wording can’t make that evidence appear. If the scope contains unrelated information, the result may emphasize something the executive didn’t ask about. You therefore inspect both sides of the mechanism. Check the report context supplied to the narrative, then check whether the generated text actually communicates the requested outcome.
Filter context is the set of filters affecting the data evaluated for a report visual or DAX expression. It can come from selections and other report conditions that narrow what the consumer is viewing. If the sales page is filtered to the western region and the current quarter, the visible charts may differ substantially from the unfiltered report. A narrative intended to explain that filtered view must be validated against that same state. Comparing it with totals from another state would tell you very little.
Take a controlled example. Begin with the sales page in its expected default state. Record a few facts visible in the relevant visuals, such as which region has the highest displayed revenue, whether margin is above or below its comparison value, and which category shows the largest visible change. Generate the narrative for the intended scope.
Then compare every material statement with those visuals. You’re looking for agreement about values, direction, category, period, and scope. Elegant phrasing isn’t the test. Agreement with observable report evidence is.
Now apply a meaningful filter, such as one region and one product category. Confirm that the visuals themselves respond as expected. Make sure the narrative has been refreshed for that filter state, then compare its claims with the newly displayed evidence.
This second test can reveal a problem hidden by the normal case. A sentence that was accurate for all regions might no longer describe the filtered page. A general statement may also be too broad for the narrow evidence now available.
A useful test changes one condition while preserving the rest of the report. Start with the default sales page and verify the narrative against the visible revenue and margin charts. Then apply only the enterprise-customer filter, confirm the charts changed, refresh the narrative, and compare again. Because the scope and prompt stayed the same, a mismatch is easier to associate with the changed report context.
If you change the filter, selected visuals, and prompt at once, a better paragraph won’t tell you which change mattered. This controlled comparison also protects against a false fix. Returning to the default state should restore a narrative supported by the original evidence, while the filtered state should remain supported by its narrower evidence.
The same discipline applies after a data refresh. First establish that the refreshed data and visuals show the expected result. Then validate the generated narrative against that refreshed state. Don’t assume that because the visual remains on the page, every sentence necessarily represents the latest condition you intend to communicate. The approved decision rule is straightforward: before using a generated narrative for executive communication, verify it against the relevant visuals and the refreshed filter state.
Suppose the analyst’s first narrative says that the central region drove revenue growth. In the default report view, the charts support that statement. When the chief operating officer filters to enterprise customers, however, the southern region becomes the strongest contributor. If the narrative still presents the original claim, the analyst shouldn’t begin by rewriting measures or changing query steps. The symptom appears in the relationship between report context and generated narrative, so that’s where the investigation begins.
Reproduce the exact consumer condition. Capture the active filters, the visuals included in the narrative’s scope, the current values shown by those visuals, and the narrative text being evaluated. Confirm that the narrative has been refreshed under the intended state. Then compare expected and actual behavior.
The expected result is not simply different wording. It’s wording that remains supportable from the chosen report evidence. Once the cause is corrected, retest the default view and the enterprise-customer view. Fixing only the failing case could create a new mismatch elsewhere.
This is why random report editing is poor troubleshooting. Changing colors, rearranging visuals, modifying an unrelated query, or rewriting a measure without evidence can conceal the original condition. Instead, follow the behavior to the layer that controls it. Query output should be checked when transformed data is suspect.
The semantic model and DAX result should be checked when values or relationships are suspect. Report interaction and filter state should be checked when the visible context is suspect. Refresh or access state should be checked when the consumer isn’t receiving the expected current experience.
Access and availability also deserve a separate check. Copilot features in Power BI depend on tenant, capacity, and model requirements. A narrative working for its author doesn’t by itself prove that the governed solution is ready for every intended consumer. If the feature is unavailable or behaves differently in the deployment context, don’t treat that as proof that the narrative’s wording or report calculations are defective. Check the relevant environment and access conditions rather than making unrelated model changes.
That distinction gives you two different kinds of failure. In the first, the analyst can’t use the intended Copilot capability because an applicable tenant, capacity, model, or access requirement isn’t satisfied. In the second, the feature produces a narrative, but the text doesn’t satisfy the business requirement or doesn’t agree with the report state.
The first calls for investigation of availability and governance conditions. The second calls for validation of scope, prompt, context, refresh state, and report evidence. Both need correction, but they don’t have the same owner.
Keep availability evidence separate from content evidence. Seeing the narrative work in one environment shows that text can be produced there under those conditions. It doesn’t establish that all tenant, capacity, model, and access requirements are satisfied everywhere the report will be used. Conversely, confirming those requirements doesn’t establish that the generated statements are correct.
One test asks whether the intended capability and consumer experience are available. The other asks whether the output is supported by the selected visuals and current state. Preserving that separation prevents an access investigation from turning into unnecessary report edits, and it prevents a readable paragraph from being mistaken for proof that deployment conditions are ready.
Human validation remains necessary even when the narrative reads naturally. Generated language can make a weak interpretation sound finished. A statement may be numerically compatible with a chart yet still be misleading because it leaves out the period, applies a conclusion too broadly, or emphasizes an irrelevant comparison. The available report context and documented limitations constrain what Copilot can produce. Treat the text as draft assistance, not as an authority that outranks the report or the analyst’s responsibility.
For executive use, inspect the meaning rather than proofreading only the grammar. Ask whether each important claim identifies the correct measure, category, period, and direction. Check that comparisons use the state visible to the executive.
Make sure the wording doesn’t imply a cause that the selected visuals merely correlate with a result. The accepted evidence says that the narrative can summarize available report content. It doesn’t justify treating generated explanation as proof of business causation.
Suppose revenue rises in the same filtered view in which delivery performance improves. A narrative may accurately mention both observations because both are available in the report context. That doesn’t show that improved delivery caused the revenue increase. The visuals may support two movements without supplying evidence of a causal relationship between them.
Human review should therefore distinguish a summary from an explanation. A supported summary names what the report shows, with the correct period, category, direction, and filter state. A stronger explanation needs evidence that the selected report content may not provide. If Copilot’s wording crosses that boundary, revise or reject the wording rather than accepting it because the sentence sounds decisive.
Usability matters alongside factual agreement. A report-wide narrative might be accurate yet too broad for a page devoted to margin exceptions. A summary of one selected visual might be accurate yet omit another visual required to understand the business tradeoff.
A prompt might request detail that overwhelms the intended audience. These aren’t reasons to abandon the feature. They’re reasons to restate the requirement and choose a scope that serves the consumer’s decision.
Return to the sales and operations report. The executives need a compact explanation of revenue and gross margin for the filters they apply during a monthly review. The analyst chooses the relevant visuals rather than unrelated report content, guides the narrative toward those measures, and validates the output in the default view. The analyst then tests a region filter, a customer category filter, and the refreshed report state. Each material sentence is compared with the displayed evidence before the report is approved for executive communication.
If a statement fails, the analyst records the expected and actual result. A wrong displayed value leads upstream to the query, model, relationship, or DAX logic that owns it. Correct visuals paired with stale or unsupported narrative wording lead to the narrative’s report context, scope, prompt, or refresh condition. Missing functionality for an intended user leads toward the relevant tenant, capacity, model, or access requirement. That path is more reliable than assuming every symptom can be fixed through one familiar setting.
This also explains why there isn’t one universally correct narrative configuration. Two reports can use the same semantic model and still need different summaries. An operational page may need commentary about delivery exceptions, while an executive page may need revenue and margin trends. A whole-page scope could suit one and distract from the other. The correct choice follows the requirement, the available context, the audience, and the governed environment.
A complete test therefore has a visible chain of evidence. The business requirement identifies what the narrative should communicate. The selected report scope identifies the evidence available to Copilot. The prompt guides the emphasis.
The visuals and active filter state provide the facts against which the output is checked. Refresh and access testing show whether the intended consumer receives the expected experience. Human review determines whether the final wording is accurate, relevant, and appropriately qualified.
When those links agree, you have more than a configuration that happened to produce text. You can explain why the narrative belongs at the report layer, what information shaped it, and how you tested its behavior under normal and changed conditions. If one link breaks, the same chain tells you where to investigate instead of sending you toward unrelated transformations, calculations, or formatting changes.
A Copilot narrative visual is useful when the requirement is written commentary about information already presented in a Power BI report. Its scope, prompt, report context, and refreshed filter state all shape what must be checked. That’s why successful text generation is only the beginning of validation. Compare each important statement with the intended visuals, test meaningful filter conditions, and investigate failures in the layer that actually owns them. When the report evidence, consumer state, and narrative agree, you can justify the result as a requirement-driven Power BI solution rather than a convenient piece of generated prose.
Review a Copilot narrative for numerical accuracy, context, and unsupported conclusions. Which source visual or measure would you use to verify each important statement? Use the episode reference to review precise terminology and the checks that are easier to inspect than to hear. This lesson was produced by Bare Metal Cyber dot com.