The Power BI Fellowship

Modern Power BI features Track

Product features layered on top of the fundamentals, most of them new to the PL-300 outline in April 2026. They don't replace a good model or correct DAX: Copilot answers better on a well-prepared model, and visual calculations are easier precisely because the model underneath is right. Each topic says when it was last checked against Microsoft's documentation, because these features change monthly.

Before you start Power BI Desktop (current monthly release) and the starter project. Copilot needs a paid Fabric capacity (F2 or higher) or Power BI Premium (P1 or higher): trial capacities and Pro/PPU licences alone aren't enough. Where you don't have it, the tasks describe what to evaluate.

Visual calculations

Data: FactSales DimProduct DimDate
Why it matters
Running totals, moving averages, percent of parent and versus-previous are the most common report calculations. Visual calculations express them in one line on the visual, without filter-context gymnastics.
Typical production failure
A model collects dozens of one-off measures ("Running Sales for Chart 3") that nobody else uses and nobody dares delete.
When to use it
Calculations that only make sense for one visual's layout: running sums along its axis, percent of the visual's parent, comparisons with the previous row.
When not to
Business definitions that other reports need (put those in model measures), or anything you need to filter, sort or export: visual calculations can't be filtered, sorted, reused across visuals or exported.

Assignments

Four visual calculations in ten minutes

A · Guided 25 min · uses FactSales, DimProduct, Dates
  1. Build a matrix with Dates[MonthName] on rows and [Net Sales] as the value (Beginner model).
  2. Select the visual → New calculation. Add Running = RUNNINGSUM([Net Sales]).
  3. Add vs Previous = [Net Sales] - PREVIOUS([Net Sales]) and Moving avg = MOVINGAVERAGE([Net Sales], 2).
  4. Add Category to rows above MonthName and add Pct of parent = DIVIDE([Net Sales], COLLAPSE([Net Sales], ROWS)).
  5. Try RUNNINGSUM([Net Sales], HIGHESTPARENT) and explain how it differs from the first running sum.
Expected result: The running sum on the March row equals the visual's total; vs Previous is blank on the first row; percent of parent sums to 100% within each category; with Reset = HIGHESTPARENT the running sum restarts for each category.

Measure or visual calculation?

C · Problem 20 min · uses —

The executive page has six calculations. Decide for each whether it should be a model measure or a visual calculation, considering reuse, filtering, export and performance.

Work out
  • Net Sales YoY % (used on four pages).
  • Running total on one line chart.
  • Percent of category total in one matrix.
  • Rank of customers used in a slicer.
  • Margin % in the KPI dictionary.
  • Difference from the first month in one table.
What good looks like: Measures: YoY %, customer rank (needs filtering/slicing), Margin % (shared definition). Visual calculations: the running total, percent of category in that matrix, difference from the first month (FIRST).

Find the limits

B · Objective 20 min · uses your matrix

Before recommending visual calculations to the team, test what they can't do.

Requirements
  • Try to filter or sort by a visual calculation.
  • Export the visual's data.
  • Copy the visual calculation to another visual.
  • Use RELATED inside one.
Expected result: No filtering or sorting by the calculation, results aren't included in data exports, it can't be reused on other visuals, and RELATED, RELATEDTABLE and USERELATIONSHIP aren't available.

Interview questions

  • What is a visual calculation?
  • What do the Axis and Reset parameters do?
  • Name some functions specific to visual calculations.
  • When should you not use a visual calculation?

Assessment

Which expression gives each row's share of its parent in a visual calculation?

  • DIVIDE([Sales], CALCULATE([Sales], ALL(Product)))
  • DIVIDE([Sales], COLLAPSE([Sales], ROWS))
  • RUNNINGSUM([Sales])
  • RELATED([Sales])

A visual calculation's results in an Export data file:

  • Are included
  • Are excluded
  • Replace the measures
  • Cause the export to fail

Rebuild one running-total measure from the Intermediate DAX topic as a visual calculation, and compare the DAX length and Performance Analyzer timing.

RUNNINGSUM([Net Sales]) vs a CALCULATE with a date filter; visual calculations operate on aggregated data, often faster.

Copilot-assisted workflows and their limits

Why it matters
Copilot can draft report pages, summarise a model, write DAX queries and suggest measure descriptions. Its answers are only as good as the model it reads, and someone still has to check them.
Typical production failure
A manager asks Copilot "what was revenue last quarter?" and gets a confident number from the wrong one of three revenue measures, because the model had no descriptions and three similar names (Sprint 05).
When to use it
Drafting pages and narratives, exploring an unfamiliar model, first drafts of DAX queries and descriptions, summaries for subscriptions. Prepare the model for AI: clear names, descriptions, synonyms, hidden helper fields.
When not to
Treating an answer as verified, using it on unprepared models for decisions, or assuming it's available everywhere: it needs paid capacity, an enabled tenant setting and a supported region.

Assignments

Prepare the model so AI answers correctly

B · Objective 40 min · uses your certified model from Sprint 05 (or the starter)

Before Copilot is switched on for the sales app, make the model unambiguous for a machine reader.

Requirements
  • Descriptions on every visible measure (from the KPI dictionary).
  • Clear names; hide keys and helper columns.
  • Synonyms for common business words (revenue, turnover, sales).
  • Decide which measure a question about "revenue" should land on.
Expected result: Every visible measure described, helper fields hidden, synonyms added, and "revenue" mapped to Net Revenue (ledger) in descriptions and synonyms, so a natural-language question has one obvious target.

Check Copilot's work

C · Problem 30 min · uses a model on F2+ or P1+ capacity

Ask Copilot five questions you already know the answers to (use Sprint 01/02 numbers), and have it draft a report page and a DAX query. Grade each output as correct, partly correct or wrong, and explain why.

Work out
  • Five questions with known answers.
  • One drafted page and one DAX query reviewed line by line.
  • What you'd tell users about trusting answers.
What good looks like: A graded table. Typical findings: right when the question maps to one well-described measure, wrong or ambiguous when names are similar or the time period is implicit; drafted DAX needs review for filters and totals. Users should verify answers against the certified report before acting.

Should we turn it on?

D · Ambiguous 20 min · uses —

Elena asks whether to enable Copilot for all report users now, for a pilot group, or not yet. Recommend one, considering capacity cost, data preparation, privacy and training.

Work out
  • Prerequisites (capacity, tenant setting, region).
  • Where Copilot compute is billed.
  • Which models are ready.
  • How you'd measure the pilot.
What good looks like: A pilot: enable for a security group on certified, prepared models only; monitor Copilot CU in the Capacity Metrics app; publish guidance on verifying answers; expand when answer quality and cost are acceptable.

Interview questions

  • What does Copilot need to work in Power BI?
  • How do you make a model better for Copilot?
  • Is Copilot output trustworthy?
  • How is Copilot usage charged?

Assessment

A user has a Power BI Pro licence and the report is in a workspace on a Pro (shared) capacity. Copilot:

  • Works
  • Isn't available: it needs F2+/P1+ capacity
  • Works in Desktop only
  • Works on trial capacity

The most effective way to improve Copilot answers about "revenue" is:

  • Longer prompts
  • Clear measure names, descriptions and synonyms in the model
  • More visuals
  • Turning off RLS

Write a half-page user guide: what Copilot is good for in our reports, how to check an answer, and who to ask when it's wrong.

Point users to the certified report and the KPI dictionary as the reference.

Paginated reports, mobile layouts, accessibility and personalization

Data: FactSales DimCustomer DimProduct
Why it matters
A report that can't be read on a phone, by a screen-reader user, or printed as a 40-page invoice list fails part of its audience, often the part that matters most.
Typical production failure
The regional sales report is used on phones in the field, but its mobile layout was never built; managers pinch and zoom, then ask for a PDF every Monday.
When to use it
Paginated reports for pixel-perfect, printable, many-page output; mobile layouts for field users; accessibility settings (alt text, tab order, contrast, titles) for everyone; personalization so users adjust visuals without copies.
When not to
Colour as the only signal, interactive reports pretending to be printouts, and paginated reports for exploration.

Assignments

Make the executive page accessible

B · Objective 40 min · uses your Beginner executive page

Run the Beginner executive page through an accessibility review and fix what you find.

Requirements
  • Alt text on every visual (what it shows, not "bar chart").
  • Tab order and titles; no information conveyed by colour alone; contrast checked.
  • Test with keyboard only and with a screen reader (Narrator or NVDA).
Expected result: Every visual has meaningful alt text (ideally dynamic, from a measure), tab order follows the reading order, status uses icons or text as well as colour, and the page can be used with the keyboard alone.

A mobile layout for field managers

B · Objective 30 min · uses your Beginner report

Regional managers check sales on their phones between store visits. Build the mobile layout for the page they use.

Requirements
  • Only the visuals that answer their question, in priority order.
  • Slicers usable with a thumb.
  • Test in the Power BI mobile app or the phone emulator.
Expected result: A portrait mobile layout with three to five visuals (headline numbers first), slicers at the top, no horizontal scrolling, tested in the mobile app.

Paginated or interactive?

C · Problem 20 min · uses —

Decide the right report type for four requests and justify each.

Work out
  • Monthly customer statements, one PDF per customer.
  • An exploratory dashboard of sales by region.
  • A 40-page price list printed for stores.
  • Commission statements emailed to each rep on the 1st (Sprint D05).
What good looks like: Paginated: statements, the price list and commission statements (fixed layout, many pages, per-recipient output, subscriptions). Interactive Power BI report: the exploratory dashboard.

Interview questions

  • When do you choose a paginated report?
  • What makes a Power BI report accessible?
  • What is report personalization?
  • Why build a mobile layout instead of relying on the phone showing the desktop page?

Assessment

Statements must be emailed as one PDF per customer each month. Best fit:

  • Interactive report with bookmarks
  • Paginated report with a data-driven subscription or parameters
  • Dashboard
  • Excel export

A KPI turns red when below target, with no other cue. Accessibility problem?

  • No
  • Yes: colour alone conveys the status
  • Only on mobile
  • Only when printed

Enable personalization on one report page and test it as a viewer.

Report settings → Personalize visuals; viewers see the personalize icon on each visual.

Analytics features: forecasting, anomalies, clustering, AI visuals and auto page refresh

Data: FactSales WebEvents DimCustomer
Why it matters
Built-in analytics answer "what's unusual, what's next, what groups exist" without a data scientist, and they're part of PL-300. They also produce confident-looking results from too little data.
Typical production failure
A forecast shaded band on three months of data goes into a board deck as a prediction.
When to use it
Forecasting and anomaly detection on line charts with enough history; clustering and grouping for segmentation; Key influencers and Decomposition tree for exploration; automatic page refresh for operational pages on DirectQuery or Direct Lake.
When not to
Presenting AI visual output as fact without checking sample size and assumptions; automatic page refresh on Import models or at intervals faster than the data changes.

Assignments

Forecast and find anomalies, then doubt them

B · Objective 30 min · uses FactSales, Dates

Add a forecast and anomaly detection to a daily Net Sales line chart of the course data, then decide whether either result is fit to show an executive.

Requirements
  • Analytics pane → Forecast; Find anomalies on the line.
  • Explain what the sensitivity and confidence settings do.
  • Judge whether 90 days of data is enough, and what you'd say on the slide.
Expected result: Both features run, but on about three months of daily data with returns and gaps the forecast band is wide and anomalies are mostly normal weekly variation. Not fit for an executive slide without much more history; say "exploratory" if shown at all.

Segment customers with clustering

B · Objective 30 min · uses FactSales, DimCustomer

Create customer segments from Net Sales and Orders with automatic clustering on a scatter chart, and compare them with a rule-based segmentation you can explain.

Requirements
  • Scatter chart by customer; ... → Automatically find clusters.
  • A rule-based grouping (e.g. ABC classes from Advanced).
  • Which would you give to Sales, and why?
Expected result: Clusters appear as a new field; they're harder to explain and can change on refresh. A rule-based segmentation (ABC by cumulative share) is stable and explainable, so it's better for Sales; clusters are good for exploration.

Real-time page, or not?

C · Problem 20 min · uses —

Operations asks for a warehouse page refreshing every 5 seconds. Decide whether and how to use automatic page refresh, considering storage mode, capacity limits and how often the data actually changes.

Work out
  • Which storage modes support it.
  • Fixed interval vs change detection.
  • Admin limits and capacity cost.
  • What interval the decision really needs.
What good looks like: Automatic page refresh works with DirectQuery (and Direct Lake) sources, not Import; admins set minimum intervals and can disable it; change detection refreshes only when a measure changes. Agree an interval with Ops based on how quickly they act (minutes, not seconds), and check capacity load.

Interview questions

  • What does anomaly detection in Power BI do?
  • Key influencers vs Decomposition tree?
  • When is automatic page refresh appropriate?
  • What's the risk of AI visuals in reports?

Assessment

Automatic page refresh works with:

  • Import only
  • DirectQuery (and Direct Lake) sources
  • Excel files
  • Paginated reports only

Key influencers finds that customers on Mobile buy less. This shows:

  • Mobile causes lower spend
  • An association worth investigating, not a cause
  • A data error
  • Nothing

Add a Decomposition tree on Net Sales by Region → Category → Product and use AI split once. Explain the result to a non-analyst in two sentences.

AI split picks the dimension with the highest (or lowest) value at the next level.