Power BI AI Visuals - Are Key Influencers and Decomposition Trees Actually Useful

Microsoft has been shipping AI-flavoured visuals into Power BI for years now. Key Influencers, Decomposition Tree, Anomaly Detection, Smart Narratives. They all show up in the demo decks and they all get a lot of marketing airtime. But once you start building reports for actual Australian businesses, you quickly find that these visuals fall into two camps. Some are genuinely useful and underused. Others are demo candy that breaks down on real data.

I've spent enough time inside Power BI tenants across mining, professional services, healthcare and financial services to have formed strong opinions on which is which. Microsoft publishes a sample report called the Artificial Intelligence Sample that walks through all of them. It is worth opening that sample if you want to see what they look like end to end, but here is what we have actually learned about each one in production.

Key Influencers - the most useful one nobody uses properly

Key Influencers is the visual where you pick a target metric, point it at a bunch of explanatory fields, and Power BI tells you which fields are driving the target up or down. In the Microsoft sample, you ask "what makes us win more deals" and it tells you that a 2% discount makes you 2.76 times more likely to win.

In our experience, this is the most genuinely useful AI visual in Power BI. It is essentially a logistic regression presented in plain English, and that is a great fit for business users who would never run a logistic regression themselves but can absolutely read a sentence that says "your customers are 3 times more likely to churn when their support tickets exceed 5 per month."

The visual is underused for two reasons. First, most analysts do not include it because it does not fit cleanly into a static dashboard layout. It looks weird next to bar charts. Second, when people do use it, they often point it at the wrong kind of data and get nonsense.

Key Influencers works when you have a target variable with enough variance and explanatory variables that have a reasonable relationship to it. It does not work when your target is essentially constant, or when your explanatory variables are themselves derived from the target. If your fields include "deal won amount" and you ask what drives "deal won status", Power BI will helpfully tell you that high deal amounts predict winning deals. That is not insight. That is data leakage.

The other thing to know is that Key Influencers needs a decent sample size. Below a few hundred rows in each category it gets unreliable. We have seen analysts run it on 80 sales records and treat the output as gospel. That is not how it works.

If you are building reports for executives in Australian financial services or healthcare, where the question "what is driving this outcome" is constantly being asked, Key Influencers is worth the investment. We embed it into the second-level pages on most of our Power BI consulting engagements. Just make sure the data scientist or analyst building it understands the underlying mechanics enough to spot when it is misbehaving.

Decomposition Tree - powerful but slow

The Decomposition Tree visual lets users break a measure down by dimensions interactively. You start with total revenue, you split by region, you split the largest region by product category, you split that by customer segment, and so on. The "AI split" option lets Power BI pick the next dimension automatically based on which one creates the biggest split.

This visual is good for ad-hoc analysis sessions. It is not good for daily reporting. The pattern we see work well is when an analyst sits down with a business stakeholder, opens the report, and uses the decomposition tree to interactively explore where a number came from. The conversation is exploratory and the visual supports that. Where it falls down is when somebody tries to build a "standard" decomposition tree that everybody looks at every morning. That misses the point of the visual.

The AI split feature is also genuinely useful, but only when your dimensions are well-curated. If your data has a dozen high-cardinality fields with no clear hierarchy, the AI split will pick whatever dimension creates the biggest difference, which is often a random employee ID or product SKU rather than a meaningful business dimension. We always recommend restricting the dimension list available to the visual to the eight or ten that make narrative sense.

The other watch-out is performance. Decomposition Tree against a large semantic model can be slow, especially with AI splits enabled. It is generating multiple DAX queries on every interaction. For models above 100 million rows, plan to either aggregate the input or accept that the visual will lag.

Anomaly Detection - hit and miss

Anomaly Detection runs a time-series analysis on a line chart and flags points that look unusual. The Microsoft sample shows it spotting an unexpected spike in software revenue on a specific day. The feature also generates "possible explanations" which try to attribute the anomaly to other dimensions in your data.

This one is hit and miss. When it works, it works well. We have had it surface anomalies in client revenue data that the finance team had not spotted, and the "possible explanations" pointed to a specific sales rep whose deal closed early and skewed the month. That is the dream scenario.

When it does not work, it generates either too many anomalies or none at all. The algorithm uses seasonal decomposition and looks for residuals beyond a configurable threshold. If your data is noisy by nature, you either suppress real anomalies or get drowned in false ones.

The other issue is that Anomaly Detection only works on line charts with a continuous date axis. It does not work on most of the visuals where you would actually want anomaly detection. So you end up building special-purpose line charts just to get the anomaly callouts, which feels backwards.

If you are looking at anomaly detection as a serious capability for, say, AI for financial services or fraud monitoring, do not rely on this feature. It is a useful talking point but it is not a serious anomaly detection engine. Use a proper tool like Azure Anomaly Detector or build something in a notebook and write the results back into your model.

Smart Narratives - the verdict is in

Smart Narratives generates a text summary of a visual or a page. "Revenue was highest in Q2 driven by software sales..." that kind of thing. When this feature first shipped a few years ago, it was clunky. The narratives sounded robotic and often missed the point.

The feature has improved since the integration with Copilot. The narratives are more natural and the underlying logic now uses the same kind of LLM grounding that Microsoft 365 Copilot uses. For executive summary pages where the user wants a quick "what should I notice" read, Smart Narratives is now genuinely useful.

The catch is that it can still hallucinate, particularly when the underlying data is unusual. We have seen narratives that confidently describe trends that do not exist in the visual. Review every published narrative carefully before it goes to a board paper.

Q&A and the future of natural language

The Microsoft sample also includes the Q&A visual where users type questions like "close % by manager" and Power BI generates a chart. This has been in the product since 2017 and it has improved steadily, but it has never been good enough to replace properly designed reports.

What is changing in 2026 is that Copilot in Power BI is starting to do this job better than the original Q&A visual. The Copilot integration uses the semantic model intelligently, understands business terms when they are defined in the model, and generates explanations alongside the chart. For most Australian organisations rolling out Copilot training for analysts and business users, the practical guidance is to lean on Copilot rather than the older Q&A visual.

Where to spend your time

If you are designing Power BI reports in 2026 and trying to figure out which AI visuals to invest in, here is the order I would prioritise.

Start with Key Influencers for any report where the question "what drives this outcome" is meaningful. Spend the time to define the explanatory fields properly.

Use Decomposition Tree for exploratory analysis pages, not standard reporting pages. Restrict the dimensions to ones that make business sense.

Use Smart Narratives sparingly on executive summary pages with manual review of the output.

Skip Anomaly Detection unless you have very clean seasonal data and you understand the algorithm. If anomaly detection is a serious requirement, build it properly outside Power BI.

The bigger trend is that Copilot in Power BI is starting to absorb a lot of what the older AI visuals tried to do. Over the next year or two I expect the standalone AI visuals to become less important and the conversational analytics experience to take over. That is a good thing for users.

Reference: Microsoft Power BI Artificial Intelligence Sample