Each page was pulled, stripped of scripts and searched for star schema, DAX, measure, fact table, relationship. The caveat cuts against me: that reads page text, not downloadable files, so a model shipped inside a .pbix would not show up.
I build dashboards for a living — the reporting our managers open every morning at a multi-branch clinic — and I package one as a template, which is my bias declared up front. It also means I've rebuilt enough "inspiring" examples to know exactly which part of them takes the time, and it isn't the part in the screenshot.
What ranks for this search — and what all of it leaves out
Google answers this query with pictures. Position zero is an image pack; position two is a Reddit thread, which is Google deciding the best available answer is people asking each other. The nine organic results break down like this: two vendor listicles with 32 and 14 examples, a custom-visuals shop claiming 2,400+ templates, a consultancy showing nine, a lead-gen blog showing ten, a six-year-old Microsoft forum thread, Microsoft's own samples index, and Dribbble — 47 design shots, many of which were never built in Power BI at all.
Now the part worth noticing. Coupler.io sells data connectors. Supermetrics sells data connectors. ZoomCharts sells custom visuals. Vidi Corp sells consulting hours. The Reporting Hub sells a trial. Every gallery on page one is a lead magnet for something that is not the dashboard.
That's not an accusation, it's an incentive structure. If you publish the star schema and the four measures behind your example, the reader no longer needs your connector or your consultant. The missing model isn't an oversight — it's the business model. Which is also why the people currently ranking are not going to fix it.
The part the screenshot hides
Every Power BI report visual quietly writes a query against a model. Microsoft puts it plainly: report visuals generate queries that "filter, group, and summarize model data," and a well-designed model is one that "provides tables for filtering and grouping, and tables for summarizing." That split has a name — dimension tables filter and group, fact tables summarize — and it's the whole game.
Microsoft's guidance is a star schema, and the wording is unusually firm for vendor docs: a star schema design "is considered a prerequisite for enterprise Power BI semantic models." The opposite move gets called out by name: writing about Dataverse specifically, Microsoft says reproducing that source schema inside Power BI "is an anti-pattern, and it's probably the toughest to overcome because it just feels right to maintain consistency." The wording is Dataverse's, but the temptation is universal.
One line matters more than the rest, and it's the one no screenshot can convey: "it's important that fact tables always load data at a consistent grain." Grain is the sentence that says what one row of your fact table means — one appointment, one invoice line, one paid claim. Get it wrong and every number downstream is wrong in a way that still looks plausible on a chart. That's the failure mode gallery examples can't warn you about, because the picture looks identical either way.
An honest caveat, since I'm quoting a vendor about its own product: all of the above is design guidance, not a measurement. Microsoft publishes no benchmark comparing a star schema against the alternatives — no dataset, no sample size. And Microsoft itself leaves the door open: "optimal model design is part science and part art. Sometimes you can break with good guidance when it makes sense to do so." Treat it as the strong recommendation of the people who wrote the engine, which is worth a lot, but is not the same thing as evidence.
"80% of the work is data prep" — where that number actually comes from
You'll meet this number the moment you start building. It is an assertion, not a measurement, and the trail is short enough to walk.
The sentence that seeded it is from a 2014 New York Times piece, and it discloses its own sourcing: "Data scientists, according to interviews and expert estimates, spend from 50 percent to 80 percent of their time mired in this more mundane labor…" Interviews and estimates — no study, no sample.
The survey usually cited behind it is CrowdFlower's 2016 report, and here the chain is worth following carefully, because it's easy to get wrong — I did, in a first draft of this paragraph. CrowdFlower's chart asks which task takes the most time: 60% of respondents said cleaning and organising, 19% said collecting data sets. Those are shares of respondents, not shares of time, so adding them isn't valid arithmetic. And CrowdFlower didn't add them — Forbes did, in the write-up that gave the number its reach; CrowdFlower's own report credits the 80% straight back to the New York Times. The sample, per Forbes, was about 80 data scientists.
The largest number I can find on this comes from Anaconda's 2021 State of Data Science: 2,030 respondents split their time as data preparation 22%, data cleansing 17% — 39% combined, Anaconda's own arithmetic — against data visualization 15% and reporting and presentation 17%.
Two honest caveats, because this number is doing work for me and I should hold it to the standard I just applied to the other one. It is self-reported — respondents typed percentages onto a fixed list — which is the same species of evidence as the NYT's estimates, only with a disclosed method and a far bigger sample. And Anaconda sells data-science tooling and surveyed its own mailing list, exactly the conflict I noted for CrowdFlower. Take it as the best-documented figure available, not as a fact.
What it still tells you is directional, and it's enough. Preparation and cleaning together outweigh anything else on the list — while the visible output, visualization plus reporting, comes to 32%. Galleries show you that visible half. The shaping, the grain and the relationships sit in the other one, and that's the part that decides whether an example survives contact with your export. (One more limit: Anaconda surveyed data scientists, not BI report builders. Reading their split onto Power BI work is my inference, not the survey's.)
The examples that do ship a model
There is one place on that first page that hands you the real thing, and it isn't a gallery. Microsoft's samples index lists 18 samples — the page calls that "some of the samples available", so treat it as a list, not a total — and most arrive as a .pbix, which Microsoft describes as containing "the report pages, data model, queries, and visuals." Open one in Desktop and you can see the relationships, the tables and the measures that produced the picture. That is the only honest kind of Power BI dashboard example: one you can take apart.
Two caveats, because I checked rather than assumed. First, the accompanying articles don't explain the model. I read the walkthrough for the Retail Analysis sample — the one Microsoft uses as its reference — and it's a click-tour of visuals: no relationship diagram, no measure explained. The model is in the file, not in the lesson. Second, the data isn't yours. A couple of samples run on a fictitious company called Contoso; most of the rest are real data that the vendor obviEnce anonymised — the Retail Analysis page says so plainly: "The company obviEnce created this sample using real, anonymized data." Either way it arrives clean and at a consistent grain, which is the one thing your own export won't be. No sample shows you duplicate patient records or a date column your system exports as text.
What it costs to actually run one
Worth knowing before you download anything. Power BI Desktop is free — Microsoft's words: "a free Windows application that lets you connect to data, transform it, and create interactive visual reports." You can open a sample, rewire it to your own export and read it on your screen without paying anyone. Note Windows: the documented requirement is Windows 10 or Windows Server 2016 or later, which is the part Mac-based practice owners find out too late.
The line is sharing. Microsoft: "You need a paid license, either Power BI Pro or Premium Per User (PPU), to publish content to a Power BI app workspace, edit content, and share content with others." A free licence can create content in My workspace; it cannot publish to app workspaces and cannot share. So: one person reading a dashboard, free. A practice manager and two partners each opening it Monday morning, paid. I worked the licensing maths out properly in Power BI Pro cost and licensing for a clinic.
How to actually read an example once you have the file
Opening a .pbix and clicking through the pages tells you almost nothing — that's just the screenshot again, with a mouse. Three things are worth opening first, in this order, and together they take about ten minutes.
The relationships view. Open the model/relationships view in Desktop. You're looking for a shape: one or two tables in the middle with lines fanning out to smaller tables around them. That's the star. If instead you see a web where everything connects to everything, or one enormous table sitting alone, you've learned something the gallery page would never have told you — and you've learned it before spending a weekend on it.
The date table. Look for a calendar table. If there isn't one, the example is leaning on Power BI's auto date/time, which is on by default and quietly builds a hidden date table behind every date column — so year-and-quarter drill-down will work, and you may never notice. It stops being enough the moment you need a fiscal year that isn't January-to-December, one shared timeline across two fact tables, or DirectQuery, where it isn't created at all. Those are exactly the things a clinic wants by month three, so an example with a real date table is one you can grow into and an example without one probably isn't.
The measures list. Expand the fields pane and read the measures before you read the visuals. A good example has few of them, named in the language of the business — Collected, No-show rate, Revenue per visit — sitting in one place rather than scattered across tables. If the numbers on the page turn out to be calculated columns rather than measures, that example won't survive being filtered in ways its author didn't anticipate. What a short, well-named measure list looks like in practice is the whole subject of what a good sample KPI dashboard actually looks like.
Do this with two or three of Microsoft's samples and the galleries stop being interesting, because you'll be able to see what they aren't showing you. If you want to build rather than read, Power BI for clinics walks an EMR export to a working report; if you are about to download a template instead, what a .pbit asks of your data shows what is inside one before you install it; the Excel-side counterpart to this article is Excel dashboard examples.
Frequently asked questions
For learning, Microsoft's official samples, because most arrive as
.pbixfiles containing the data model rather than as screenshots. For visual inspiration the vendor galleries are fine — just know that eight of the nine pages ranking for this query show you a picture and explain no model. Microsoft's index is the exception, and it hands you the file instead.Microsoft's samples index publishes direct downloads: most come as a
.pbix, eight of them also as an.xlsx, and a couple are one format only — COVID-19 US is a.pbittemplate, and the Financial sample is a workbook rather than a report. That's still the reliable route. Most gallery "templates" either sit behind an account signup or need the publisher's own connector product.Not to use it yourself. Power BI Desktop is free and opens any
.pbix. You need Pro or Premium Per User to publish to an app workspace and to share with other people — so the cost starts when a second person needs to open it.Almost always the model, not the visuals. Samples are built on clean, invented data at a consistent grain; real exports arrive with duplicates, text-formatted dates and a different grain than the one the measures assume. Microsoft's own guidance stresses that fact tables must load at a consistent grain — that's the sentence doing the work.
Microsoft calls it a prerequisite for enterprise semantic models, and — writing about Dataverse specifically — calls reproducing a source system's schema inside Power BI an anti-pattern. But it also says model design is "part science and part art" and that you can sometimes break with the guidance. It's a strong recommendation from the people who wrote the engine, not a benchmarked result. For a small clinic report, a clean fact table at one grain plus a real date table gets you most of the benefit.
WRITTEN BY
Olha · clinic data analyst
I build the reporting our managers open every morning at a multi-branch medical clinic — and package it so other practices don't have to start from scratch.
This draft went through an adversarial review briefed to refute it, and six things changed. The opening claim read "not one of the nine" until the review pointed out that Microsoft's samples index is itself one of the nine and does ship the model — it now reads eight of nine, and the method's blind spot (a text search cannot see a model inside a downloadable file) is stated in the lede rather than buried. The paragraph attacking the "80%" figure credited the 60+19 addition to CrowdFlower; it was Forbes that added them, CrowdFlower credits the New York Times, and those two percentages are shares of respondents rather than shares of time — so this article had committed the exact arithmetic error it was prosecuting, and now says so in the text. The Anaconda figure was called "measured"; it is self-reported, self-selected and vendor-run, and is labelled that way now, because applying that disqualifier only to the number I disagreed with was precisely the asymmetry the review was hunting. The claim that sample data is "invented" was wrong and contradicted the Microsoft page cited beside it — most samples are real data anonymised by obviEnce. A statement that Microsoft ships both file formats for all 18 samples was false, and it had reached the FAQ schema where answer engines would have quoted it. Finally the date-table paragraph was technically wrong: Power BI's auto date/time is on by default and covers the simple cases, so the honest version explains where it actually stops instead. One uncited sentence about a 2007 ETL paper was cut rather than softened.