Skip to content
Tutorials Center
My learning

Topics

Lesson 42 of 56 ยท Visuals and Reports

AI Visuals: Key Influencers, Decomposition Tree and Smart Narrative

Use Power BI's AI visuals: key influencers, decomposition tree, smart narrative and the Q&A visual. Learn how they work, what to feed them and when not to trust them.

Most visuals answer a question you already know how to ask: sales by region, orders by month. AI visuals help with questions you do not yet know how to phrase: what drives customers to leave, where does this number come from, what should a manager read first. They are built into Power BI Desktop and the service, they need no code, and they all work on the data in your model.

In this lesson you meet four of them: Key influencers, Decomposition tree, Smart narrative and the Q&A visual. You will learn what to put in each field well, how to read the results, and why none of them proves a cause. The visuals are point-and-click, so the lesson describes the steps. The runnable examples are Python stand-ins that compute the simple statistics behind the same ideas on a small synthetic churn table. They are explanations, not Power BI's own algorithms.

Which AI features you see depends on your Power BI version, your license and your tenant settings. Copilot-based options, and the Q&A experience in particular, also require specific capacities and admin settings. Treat this lesson as the foundation and check what your organization has switched on.

A synthetic churn table

All stand-ins below use the same table: 600 customers, each with a contract type, a number of support calls last quarter, a region and a Churned flag. The data is generated with a fixed seed, so you get the same numbers every time. By construction, monthly contracts and many support calls raise the chance of leaving, and region does nothing. A good tool should find the first two and ignore the third.

Key influencers

The Key influencers visual answers "what influences this outcome?". You give it:

  • Analyze: the metric to study. It can be a category ("Churned = Yes") or a numeric column, such as a satisfaction score.
  • Explain by: the fields that might influence it (contract, support calls, region, tenure, plan...).
  • Expand by (for a measure-like analysis): the unit of analysis, such as the customer ID, when the analyzed value is a measure.

It has two tabs. Key influencers lists each factor with a statement such as "When Contract is Monthly, the likelihood of Churn increases by 1.4x", and a bar showing the size of the effect. Top segments groups customers into segments (for example monthly contract and 3+ support calls) ranked by how much their churn rate differs from the average, and shows the size of each segment.

Behind the scenes it runs a regression: logistic regression for a category and linear regression for a number, then ranks the fields by how strongly they shift the result. It handles a mix of numeric and categorical fields and the effect sizes are comparisons, not causes.

Try it: lift by field

A simple way to rank factors is lift: a group's churn rate divided by the overall churn rate. A lift of 1.5 means "50 percent more likely than average". Power BI's regression takes more into account (for example it separates the effect of two correlated fields), but lift is the idea you read in the visual.

Python Stand-in: churn rate and lift for each value

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Contract and support calls spread the lift widely. Region barely moves it, so it ranks last: that is what an uninformative field looks like in the Key influencers tab, where it either does not appear or has a very small effect.

What makes the visual trustworthy

  • Enough rows. A rule of thumb from Microsoft's guidance is at least a few hundred rows, and at least about ten observations for each category you want a conclusion about. Small groups produce extreme, unstable percentages.
  • Relevant fields. Remove IDs, names and fields that are consequences of the outcome (a Cancellation date column "explains" churn perfectly and tells you nothing).
  • Numeric fields are treated as continuous by default, and you can set them to categorical if their values are codes.
  • Correlation, not causation. Customers with many support calls leave more, but the calls may be a symptom of a problem rather than the cause.

Decomposition tree

The decomposition tree lets you drill into a total one field at a time. You give it a measure to Analyze (for example Churn Rate or Total Sales) and the fields to Explain by. It starts with one bar for the total, and you click the + to split it by a field of your choice, then split again.

The smart part is the AI splits: instead of choosing the field, you choose High value or Low value, and Power BI picks the field that gives the highest (or lowest) value of the measure in one of its members. This is how you find the "hot spot" quickly: the highest churn rate is among customers with 4 support calls, and within them among those on a monthly contract. You can lock a level so the tree keeps it when you change other things.

Try it: a greedy "high value" split

Python Stand-in: a decomposition tree that follows the highest churn rate

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Each level follows the single highest bar, which is a greedy search: it finds a path to a high value, not the best possible segment. The tree is best for exploring and explaining a number to colleagues, because you can click through it in a meeting. Use Key influencers when you want a ranked list of drivers.

Smart narrative

The Smart narrative visual writes a text summary of the page or of a selected visual: the highest and lowest values, the trend, a comparison with the previous period. It updates automatically when filters change, so the summary always matches the numbers.

You can edit the text, add your own commentary and insert dynamic values with the + value button: pick a measure or a field with a conditional format, and the number in your sentence stays in sync with the model. For a page you re-use each month, a narrative with two or three dynamic sentences ("Revenue was {Revenue}, {change} versus last month") saves writing the commentary by hand.

Python Stand-in: a narrative that follows the data

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The sentences are generated from the numbers, which is the whole idea. Always read the generated text: a summary that says "sales increased" is still wrong if the filter you set is a single day.

The Q&A visual

The Q&A visual lets a reader type a question such as "total sales by region last year" and gets an answer as a chart. It reads the tables, columns and measures of your model, so naming is everything: Total Sales works, Measure1 does not. You improve it in the model with:

  • Synonyms (in Model view): "revenue", "turnover" and "income" can all point to Total Sales.
  • Suggested questions that you pre-set for readers.
  • Hidden or renamed fields that you do not want in answers (hidden columns are not used).
  • The Q&A setup (Linguistic schema) for teaching the engine phrases and relationships.

You can add a Q&A visual to a report page, or a Q&A box to a service dashboard (see Power BI Service). In newer versions, Copilot adds natural-language report building on top of the same model quality, with the same rule: a model with clear names, a star schema and good descriptions gives better answers.

When to trust AI visuals

Check Question to ask
Sample size How many rows are behind this statement?
Direction Could the "cause" be a result of the outcome?
Missing fields Is an important factor not in the model at all?
Stability Does the answer stay the same if I filter to the last year?
Domain sense Would a person who knows the business find it plausible?
Python Stand-in: small groups produce extreme percentages

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The highest rates come from cells with four or five customers, where 100 percent or 80 percent is close to meaningless. Requiring a minimum size gives results you can act on. The key influencers visual applies similar safeguards, but you should still check the size of any segment before telling anyone about it.

Practical advice

  • Prepare the model first: clear names, a star schema, hidden helper columns.
  • Choose Explain by fields that a business person can influence, not IDs and not consequences.
  • Use Key influencers to rank drivers, the decomposition tree to explain a number, Smart narrative for commentary and Q&A for self-service questions.
  • Validate a surprising finding with a normal chart before you present it.

Check your understanding

Quick check ยท 5 questions

of 5 answered You scored / 5 ยท perfect!

  1. 1. What does the Key influencers visual's Analyze field hold?

    Analyze is the outcome, such as Churned = Yes. Explain by holds the candidate drivers.
  2. 2. What does a lift of 1.4x for monthly contracts mean?

    Lift compares the group's rate with the overall rate. It describes association and does not prove a cause.
  3. 3. What is the difference between 'High value' and 'Low value' AI splits in the decomposition tree?

    An AI split picks the field that leads to the highest or lowest value of the analyzed measure, so you find the hot spot without trying each field.
  4. 4. How can you make the Q&A visual understand 'revenue' as your Total Sales measure?

    Synonyms in Model view teach Q&A alternative words for tables, columns and measures.
  5. 5. Which is a reason to distrust a key influencer?

    Very small groups produce extreme percentages by chance. Check the size of the segment.

Summary

  • Key influencers ranks the drivers of an outcome and lists top segments; Decomposition tree drills into a measure with manual or AI splits; Smart narrative writes live text; Q&A answers typed questions.
  • They need a clean model with clear names, enough rows and relevant fields.
  • Results show association, not causation. Check segment sizes and confirm surprises with a normal visual.
  • Availability, especially for Copilot features, depends on version, license and admin settings.

Next: Custom Visuals, AppSource, and R/Python Scripts.