A dashboard nobody opens is not a design failure. It is a specification failure — it was built from the data that happened to exist rather than from the decision it was supposed to serve.
Most large organizations own a graveyard of these. They were requested, scoped, built competently, demonstrated to appreciative stakeholders, and then opened four times in the following year. The build quality is rarely the issue. The specification was wrong before anyone touched a tool.
From the standard
“Power BI semantic models represent a source of data that's ready for reporting and visualization.”
The inversion that causes it
The usual sequence starts with availability. Someone asks what data we have, an analyst enumerates the warehouse, and the dashboard becomes a tour of it. Every table that could be visualised is visualised, because leaving something out feels like an omission.
The result is comprehensive and useless. It answers no question in particular because it was not built to answer one, and a reader who cannot find what they need in ten seconds goes back to asking a person.
The correct sequence runs the other way: start from a decision, and let it determine what data is required — including, importantly, data you do not currently have.
The specification question
Before anything is built, get a concrete answer to: what decision does someone make on a recurring basis, who makes it, how often, and what would have to be true for them to decide differently? If the requester cannot answer that, the dashboard is not ready to be specified — and building it anyway is how the graveyard grows.
Three failure modes at enterprise scale
Nobody owns the decision
A dashboard commissioned for "the leadership team" belongs to nobody. Attendance at a weekly meeting is not ownership, and a page without a named reader accumulates requests from everyone present until it serves none of them.
Name one person. Others may read it; one person decides what belongs on it.
The metric is not actionable at that level
An executive shown a number they cannot influence has been informed, not equipped. Total company revenue on a regional manager's page is context; it is not a decision input, and it dilutes the numbers that are.
Every metric on a page should pass a simple test: if this moved, would this reader do something different? If not, it is decoration.
The number is not trusted
This one kills dashboards silently and permanently. A reader spots a figure that disagrees with the spreadsheet they have maintained for three years. Nobody can immediately explain the discrepancy. They return to the spreadsheet, and the dashboard is dead — not because it was wrong, but because it could not defend itself.
Which is why the semantic layer matters more than the visuals. A metric with one governed definition, inspectable and versioned, can be reconciled in a meeting. A metric computed inside a report cannot.
Build the first version faster than you can argue about it
Requirements for a dashboard are famously unreliable because people cannot specify what they want until they see something wrong. Three rounds of written requirements produce less clarity than one draft in front of the reader.
This is the strongest argument for automating first-draft generation, and it has nothing to do with headcount. When a draft takes ninety seconds rather than three days, the conversation changes from negotiating a specification in the abstract to reacting to something concrete — which is the only mode in which most stakeholders can be precise.
The condition is that the draft must be a real artifact in your own platform — a project file with the semantic model, measures, and layouts, which your team can version and extend. A screenshot or a hosted preview does not survive the second iteration, because the moment the reader asks for a change, someone has to rebuild it properly anyway.
Retire pages deliberately
Dashboards accumulate and are almost never removed, because removal requires someone to assert that nobody needs a thing. So the estate grows, the useful pages become harder to find among the abandoned ones, and new readers cannot tell which is authoritative.
Instrument usage and review it quarterly. A page with no views in ninety days should be archived by default, with the owner required to object rather than to approve. The estate stays legible, and the surviving pages are the ones somebody actually decides with.
The ten-second test
Show the page to its named reader for ten seconds, then take it away and ask what state the business is in and what they would do about it. If they cannot answer, the problem is not that data is missing — it is that too much is present and nothing is prominent.
Axionalytics
Production agentic AI for enterprise engineering, data, and revenue teams.