Analytics and AI
- Configure Q&A so natural-language questions resolve to the right measures
- Read a key influencers visual as correlation evidence — and say so out loud
- Decompose a measure interactively with the decomposition tree
- Apply forecasting and anomaly detection with their assumptions stated
- Use Copilot the way this course has taught you to use everything: verify before you trust
Estimated time: 60–75 minutes. Feature availability varies by tenant and license — the principles are examinable everywhere; the clicks depend on your lane.
Why this matters
The last lesson of this course is about the features that answer questions for you — natural language, machine-found drivers, generated pages, generated DAX. They are genuinely useful, and they share one property with every silent failure this course has cataloged since the alphabetical Aprils of Lesson 2: they produce confident output whether or not it is right. A truncated axis at least sits still while you inspect it. A generated answer smiles back. The discipline you have practiced for thirteen weeks — predict, verify, name what a number is measured against — is not made obsolete by AI features. It is what makes you allowed to use them.
Q&A: the model answers in English
The Q&A visual takes typed questions — "total revenue by region in 2025" — and builds the visual it thinks you asked for. Under the hood it maps your words onto the model you built: table names, column names, and above all measure names. Which lands the L05 rule in a new place: a model whose numbers live in explicit, well-named measures answers questions correctly; a model of implicit sums answers with guesses. Q&A is your naming discipline, audited by a robot.
The craft: test the questions readers will actually type; add synonyms (Modeling → Q&A setup) so "sales" and "income" resolve to [Total Revenue] instead of a half-right implicit sum; and review Q&A's log of unanswered questions — like the export button, it is telemetry about what readers need. The check that never lapses: when Q&A renders an answer, read which fields it chose (shown under the visual) before trusting the number. "Average revenue" is still two questions — Lesson 5's ambiguity does not go away because a parser picked one silently.
Key influencers: correlation, labeled honestly
The key influencers visual takes a target — what drives Return Rate? — and ranks fields by statistical association. Run it on Lantern & Pine's returns and it will find… not much, because you already know the truth: return rates run a narrow 2.44–3.82% band with no strong driver. That null result is the honest reading, and a student who reports "the influencers visual found weak effects, consistent with the uniformity we measured in Lesson 8" has understood both the tool and the data.
Two disciplines when it does find something. First, the visual reports correlation — "returns are higher when category is Furniture" is an association, not a mechanism, and the word "drives" in your write-up is a claim your evidence does not license. Say "is associated with." Second, check the base sizes it reports: an impressive ratio on forty rows is the base-effect lesson from L06 wearing lab goggles.
The decomposition tree: drill-down, freed
The decomposition tree shows a measure — [Total Revenue], say — and lets the reader split it by any dimension, repeatedly: company → region → category → subcategory, in any order, interactively. It is L09's drill-down without a fixed hierarchy, and its honest use is exploration: "where does the Midwest's revenue live?" Its "AI splits" option (high/low value auto-selection) picks the next split for you — useful, and worth the same glance you give any automatic choice: which field did it pick, and would a different split tell a different story?
Forecasting and anomaly detection: assumptions on display
Add a forecast to the monthly revenue line and Power BI extends it with a confidence band. Two honest notes. First, the forecast assumes the future resembles the past's pattern — trend plus seasonality. Our data forecasts beautifully because it was generated with stable growth and clean seasonality; real revenue meets pandemics, competitors, and category launches the band knows nothing about. A forecast is a pattern-extension, not a promise, and presenting one without its band (or its assumption) is the truncated axis of the time dimension. Second, anomaly detection flags points outside expected ranges — and every flag deserves the L03 investigation before action: May's file was "anomalous" too, until seasonality explained it. Detection is a doorbell, not a verdict.
Copilot: the intern with infinite confidence
Copilot in Power BI drafts report pages, writes DAX from descriptions, and summarizes visuals in prose. Used well, it is a fast first draft of things you know how to verify. The three uses and their checks:
| Use | What it does well | Your check |
|---|---|---|
| Draft a page | Reasonable starting layouts from your model | The L10 design pass — question test, hierarchy, titles-as-claims. Drafts label nothing correctly by default |
| Write DAX | Syntax, boilerplate, pattern recall | Read it like L02's inherited recipe: what does each part do? Then verify the result against a figure you know — your manifest habit |
| Summarize | Fluent prose about what a page shows | Every number in the prose checked against the visual. Fluent ≠ faithful, and a fabricated figure reads identically to a real one |
Where this goes next
Three roads from here. PL-300 — Microsoft's Power BI Data Analyst certification — covers what this course taught, with an exam-shaped emphasis; a focused review of Modules 1–2 plus practice exams is the remaining distance. Fabric is the platform around everything you touched — dataflows, warehouses, pipelines — the data-engineering side of the star schemas you consumed. And the third road is the one this course was really about: being the person in the room who asks "average of what?", "compared against which base?", "what does this exclude?" — in any tool, against any confident number, for the rest of your career.
Common mistakes
- Shipping Q&A without synonyms or testing. Readers type "sales"; the model says [Total Revenue]. Bridge it, or the robot guesses.
- "Drives" in a write-up of a correlational visual. Associated with. The mechanism is your hypothesis, not the tool's finding.
- Forecasts without bands or assumptions. A line extended confidently into the future is a claim; the band and the "assumes past patterns" sentence are its honesty equipment.
- Acting on anomaly flags before investigating. Doorbell, not verdict — May's file taught this before the AI did.
- Trusting generated output because it rendered. The course in one sentence, inverted. Verify against a figure you know.
Summary
- Q&A audits your naming: explicit measures + synonyms = right answers; read the field choices before quoting.
- Key influencers reports association — and a null result on uniform data is a finding, not a failure.
- The decomposition tree is freeform drill-down; glance at what its AI splits chose for you.
- Forecasts extend patterns with stated assumptions; anomalies ring doorbells that investigation answers.
- Copilot drafts; you verify — against known figures, with the same discipline that caught every silent failure this course planted. That discipline was the course.