Using Power BI Insights to Explain Why the Numbers Moved
The most common question anyone asks a dashboard is not "what is the number." It is "why did it change." Revenue dropped 8% last month and the CFO wants to know what drove it. Support tickets spiked on Tuesday and the ops manager wants the reason before the 9am stand-up. A region is suddenly underperforming and nobody can say whether it is one bad account or a broad slide. The chart shows the what perfectly well. The why is where people spend their afternoon slicing and dicing.
Power BI has a set of features aimed squarely at that question, filed under Insights. Most teams either do not know they exist or tried them once, got an underwhelming result, and never went back. Both are a shame, because used well they save real time, and used badly they mislead. Microsoft's overview is Find insights in your reports, and it is a fair introduction. This post is the consultant's version - what we actually get out of these tools on client work, and where we tell people not to trust them.
What is actually in the box
"Insights" is not one feature, it is a small family of them, and it helps to keep them straight.
Quick Insights is the oldest. Point it at a dataset in the Power BI service and it runs a batch of algorithms across your data looking for patterns - outliers, trends, correlations, seasonality, majority factors - and hands you back a scrollable set of auto-generated visuals with a one-line explanation on each. It is the "show me anything interesting" button. You can run it on a whole dataset or, more usefully, on a single tile to get insights focused on that one result.
Explain the increase / decrease is the one people love once they find it. Right-click a data point on a line or bar chart - say the month where revenue fell - and Power BI analyses which categories in your model contributed most to that movement. It comes back with "this drop was mostly driven by these two product lines in these three regions" and offers you a few different chart shapes to view the breakdown, which you can then drop straight onto the canvas. This is the feature that answers the CFO's question in about four clicks instead of forty.
Find where the distribution is different is the quieter sibling and honestly the one I rate most highly. It looks at how a measure is distributed across a dimension and finds where that distribution breaks the pattern - where one segment behaves unlike the rest. It is good at surfacing the "one weird region" problem that a headline average hides completely.
All of this is AI-assisted analysis running over your model, not a chatbot. You are not typing questions. You are right-clicking a point and asking the engine to go find the contributors for you.
Where it genuinely earns its place
The honest win is speed on the first pass of a "why" question. When a client rings up mid-month asking what is behind a swing, the old workflow was to open the model, build a couple of breakdown visuals, cross-filter by region, then by product, then by segment, and slowly triangulate. Explain the increase does the first cut of that in seconds. It is very good at the obvious-in-hindsight answer - the kind where, once you see it, you think "of course, it was the two enterprise accounts that churned." That is most business questions, most of the time.
It is also a decent teacher for people who are not analysts. A regional manager who would never build a decomposition tree can right-click a dip and get a plain breakdown of what moved. That democratising effect is real. We have watched non-technical staff start answering their own "why" questions instead of queuing up a request with the BI team, and that is a genuine productivity gain, not a slide-deck one.
Quick Insights is more of a discovery aid. I would not build anything on it, but running it over a fresh dataset early in a project sometimes surfaces a correlation or an outlier we would have got to eventually, faster. Think of it as a second pair of eyes that never gets bored, not as an analyst.
Where it falls down, and this matters
Now the honest part, because these tools have failure modes and they are the dangerous kind - the kind that produce a confident, plausible answer that is wrong.
The big one is that correlation gets dressed up as explanation. "Explain the increase" tells you which categories in your model are associated with the movement. It does not know causation from a hole in the ground. If ice cream sales and sunburn both rise, it will happily attribute one to the other if they share a dimension. The feature finds where the numbers moved together. Whether that is the cause, a symptom, or a coincidence is a judgement it cannot make and does not pretend to. People forget that and quote the output as if it were a root cause. It is a hypothesis, not a verdict.
The second problem is that it is only as good as your model. If your data model is missing the dimension that actually explains the change - say the real driver was a pricing change that lives in a system Power BI cannot see - then Insights will confidently attribute the movement to whatever dimensions it does have. It cannot flag "the real cause is not in this dataset." It just works with what it has and sounds equally sure either way. A well-shaped star schema with the right descriptive attributes gets dramatically better results than a flat, half-modelled table, which is one more reason model quality is never optional. This is a lot of what we tighten up during a Power BI consulting engagement before we let anyone lean on features like this.
Third, the quality is uneven. On clean, well-structured data it is impressive. On messy data with high cardinality, lots of nulls, or weird distributions, the output can be noise - a scroll of visuals that are technically true and completely useless. There is no shortcut here. Good inputs, useful insights. Rubbish inputs, rubbish insights, delivered with the same confident tone.
How we tell clients to use it
Our rule of thumb is simple: use Insights to find the question, then verify the answer yourself. It is a fast way to generate a hypothesis about what drove a change. It is not a substitute for the ten minutes of human checking that confirms the hypothesis holds up. Right-click, get the suggested drivers, then actually look at those drivers, sanity-check them against what you know about the business, and confirm the story makes sense. The tool narrows the search space from everything to a handful of candidates. You still do the thinking.
We also tell people to treat it as a private exploration tool, not a reporting one. It is brilliant for the analyst poking at a number before a meeting. It is risky as a self-serve button handed to an audience who will take the output at face value and repeat it upstream as fact. The output needs a literate reader. Know your audience before you point them at it.
Where this is heading
Insights is the older, menu-driven face of a much bigger shift. The natural-language and Copilot experiences now landing in Power BI and Fabric are pushing the same idea further - you will increasingly just ask "why did revenue drop in Queensland last quarter" in plain English and get a narrated answer with the breakdown attached. The Insights features are the foundation that grew into that, and the same strengths and cautions carry straight over. Faster to a hypothesis. Still no substitute for judgement. Still only as trustworthy as the model underneath.
That last point is the one worth ending on, because it is the same in every generation of this tooling. Whether it is a right-click menu from a few years ago or an agent answering in a sentence, the AI is reasoning over your semantic model. If the model is clean, well-related and carries the attributes that actually explain your business, the answers are genuinely good. If it is not, you get fast, fluent, confident nonsense. Getting that foundation right is most of what we do across AI for business intelligence and the Microsoft Fabric work that increasingly sits underneath it.
So go turn Insights on and use it. Right-click the dip, read the suggested drivers, and let it save you the first half hour of digging. Just remember that it found you a lead, not a conviction, and the checking is still your job.