A 100% perfect report can still be useless to the organisation.

The data is validated, the DAX measures are correct, the summary was even generated by AI. And yet no one can make any decisions from it. The report is not poorly built: it has precisely answered a question that no one was, in fact, asking. Recent Power BI productivity has solved one problem and created another. Copilot already generates automatic summaries, suggests DAX measures and identifies anomalies without constant human intervention. This means that building a technically correct report is no longer the main point of friction in the process. The problem now is different: ensuring that this technically flawless report answers the right question.

The figures confirm the speed of this shift. Gartner predicts that 75% of all new analytical content will be contextualised by generative AI by 2027, and that by 2028 AI-generated narratives and dynamic visualisations will replace 60% of traditional dashboards. This is not a distant forecast; it is the planning horizon for any BI team designing reports today.

AI will not generate more bad dashboards. It will generate more correct and irrelevant dashboards, faster than any team can review. It may seem like a subtle distinction, but it is what separates a BI team that adds value from one that merely produces reports.

Three problems, not one

It is easy to conflate three different discussions when talking about dashboards that do not deliver: productivity driven by AI, information design, and governance over who decides what is left out. These are distinct problems with distinct solutions. Solving only one of them is not enough, which is why this article addresses them in the right order.

The symptom we already knew

There is a situation that repeats itself in almost every organisation we work with. Someone requests a report; the BI team delivers a Power BI report with fifteen visuals and six pages, with all KPIs calculated. Three weeks later, no one can explain what happened in the business over the last six months without opening Excel alongside it, “just to double-check”.

It is always the same picture: a dense matrix of numbers where a single card with a trend indicator would solve the problem in seconds, or a grid of KPIs where everything competes for attention at the same time and nothing truly stands out. The problem is rarely technical: the semantic model is well built; the DAX measures are correct. The dashboard was designed to contain information, not to answer a specific question.

The design problem: containing information vs answering a question

The clearest example is also the most common. Take the question any Sales Director asks at the beginning of each semester: where is the company losing revenue this semester? A “correct” report may contain all the data needed to answer this. And yet still force the reader to search for the answer across fourteen visuals and five fixed filters at the top.

How most reports respond: everything visible, nothing highlighted. The answer is somewhere in there./em>

The same data model, the same source, but designed around the question rather than around the available data, produces a very different outcome:

How it should be: the answer first, the context second, the recommended action clearly visible.

In this specific case, the answer is immediate: the South region accounts for 69% of the revenue decline in the semester, an impact of -€443,809 compared to the previous period. The difference between the two reports is not in the volume of data or the technical quality of the model. It is the same semantic model in both cases. The difference lies in someone having the authority to decide that this question mattered more than the other forty that could also have been asked. Which brings us to the second problem.

The institutional problem: why this keeps happening, even without AI

This is the part that most articles on storytelling with data tend to ignore: the usual cause is not a lack of design knowledge. It is institutional fear.

In the report audits we carry out, the pattern repeats itself regularly: a dashboard grows visual by visual, page by page. Rarely because someone decided they needed fifteen charts. Almost always because, meeting after meeting, someone asks, “shouldn’t this be included as well?” and no one has the authority, or the willingness, to say no. The result is a report built through defensive accumulation, where each element exists “just in case” rather than to support a specific decision.

This explains a common phenomenon: two people leave the same meeting, look at the same dashboard, and arrive at two different interpretations of what “went wrong” that month. Not by chance—a dashboard built by accumulation, without anyone deciding what to leave out, does not tell a story. It tells as many stories as there are people who asked to add “just one more thing”.

This is not solved with colour or typography. It is solved with governance: defining who, within the organisation, has the authority to decide what a report will not show. Without that definition, any design improvement lasts until the next meeting.

Recent studies help explain the persistence of this fear: 66% of marketing and sales professionals feel anxious when working with data. An overloaded dashboard is not born solely from the defensive accumulation of those who build it. It is often the response to an audience that does not trust itself to interpret less information and prefers to ask for more rather than risk not having the answer at hand when needed.

Depth is hidden, not removed

The most common objection to this type of dashboard is always the same: “But what if I want more detail?”. There is no need to sacrifice the focus of the main page to address that. You simply provide a second layer, accessible with a click, without forcing it on those who only need the answer.

Drill-through page: the same data model, the same question, now at customer and product level, visible only to those who need to go deeper.

By clicking on the South region bar, the analyst is taken to a page automatically filtered for that region: the specific customers behind the decline, the most affected products, and the monthly trend across the semester. It follows the same “one question, one answer” logic applied in cascade: the first page answers “where”, the second answers “who” and “what”. Neither forces the other to become more cluttered.

What AI is changing (and what still requires human judgement)

The May 2026 update introduced the “Summarise” button in the header of each visual: a single click generates a summary of what changed and what stands out. In June, Copilot in web modelling was introduced, capable of suggesting DAX measures from natural language. This is a genuine productivity gain.

However, we have already seen teams fall into a trap: treating that summary as the final product, without reviewing which insight should lead the narrative. AI compresses the distance between raw data and readable insight, but it does not decide which insight matters most. That remains a human decision, made by someone with the authority to make it—the same logic behind the institutional problem described earlier, now applied to a new tool. AI removes the mechanical effort from the equation. The responsibility for deciding what to leave out still belongs to people.

Practical recommendations

Five actions that address, in order of the three levels described above, the institutional problem, the design problem, and the role of AI:

1. Assign explicit authority to say “no” – Before solving visual hierarchy, you must define who can refuse a request for “just one more KPI”. Without that authority, even the best design will not survive a second meeting.
2. Write the question the report answers before opening Power BI Desktop – If the question cannot fit into a single sentence, the report is not ready to be built.
3. Give depth to those who ask for it, not to everyone by default – A well-designed drill-through solves this without forcing the main dashboard to grow.
4. Use Copilot to accelerate, never to decide – Let it generate the first summary or the first DAX measure, but always review which insight is leading before publishing.
5. Test the reading flow, not just the numbers – Ask someone outside the team to look at the dashboard for 30 seconds and state, from memory, the main conclusion. If the answer does not match the intended message, the problem is not in the data. It lies in the hierarchy—or in someone, somewhere in the process, being reluctant to enforce it.

In the end, the perfect dashboard is not measured by the volume of correct data, nor by the quality of an AI-generated summary. It is measured by the presence of someone with the authority to decide what mattered—and the discipline to exclude everything else.

Sources:

• Gartner (2025): https://www.gartner.com/en/newsroom/press-releases/2025-06-18-gartner-predicts-75-percent-of-analytics-content-to-use-genai-for-enhanced-contextual-intelligence-by-2027
• Sisense (n.d.): https://www.sisense.com/reports/gartner-ai-powered-analytics/
• Marketing LTB (n.d.): https://marketingltb.com/blog/statistics/storytelling-statistics/
• Fuselab Creative (2025): https://fuselabcreative.com/top-dashboard-design-trends-2025/
• Fuselab Creative (2026): https://fuselabcreative.com/top-data-visualization-trends-2026/
• Indiana Wesleyan University (2026): https://www.indwes.edu/articles/2026/01/data-storytelling-managers-dashboards-into-decisions
• USDSI (2026): https://www.usdsi.org/data-science-insights/what-powers-data-storytelling-in-2026
• Microsoft Docs (n.d.): https://learn.microsoft.com/pt-br/power-bi/create-reports/copilot-introduction