Lecture video · L12 Analytics & AI

[Slide 1] The last lesson of this course is about the features that answer questions for you: natural language, machine-found drivers, generated pages, generated formulas. They are genuinely useful. And they share one property with every silent failure this course has cataloged since the alphabetical Aprils of week two: 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 is not made obsolete by these features — it is what makes you allowed to use them. Tonight we go through each one: what it does well, how it fails, and how you check.

[Slide 2] The Q and A visual takes typed English — total revenue by region in twenty twenty-five — and builds the visual it thinks you asked for, by mapping your words onto the model you built: table names, column names, and above all measure names. Which lands the Lesson Five rule somewhere new: a model of explicit, well-named measures answers questions correctly, and a model of implicit sums answers with guesses. The craft is synonyms — bridging the words readers actually type to the measures you actually named — and testing the questions your stakeholders will ask before they ask them. And one check never lapses: when Q and A renders an answer, read which fields it chose, shown right under the visual. Average revenue is still two numbers nearly double apart, and the parser picks one without telling you. Lesson Five's ambiguity survives natural language completely intact.

[Slide 3] The key influencers visual takes a target — what is associated with high return rates — and ranks fields by statistical association. Run it on our returns and it finds almost nothing, and that is the honest result: you measured the uniformity yourself in Lesson Eight, a band from two point four four to three point eight two percent with no crisis cell. Two instruments agreeing on no strong driver is a finding, not a failure — and rerunning until drama appears is the stretched color scale wearing statistics. When it does find something: the visual reports correlation, so your write-up says associated with, never drives — the mechanism is your hypothesis, not the tool's output — and you check the base sizes it reports, because an impressive ratio on forty rows is February's base effect in lab goggles. The decomposition tree, meanwhile, is drill-down freed from fixed hierarchies: split revenue by anything, in any order, live. Its AI-split option picks the next dimension for you — useful, and worth the same one-second glance you give every automatic choice.

[Slide 4] Prediction beat, and it is the whole lesson in one scenario. Copilot writes you a year-over-year measure. It renders without a single error, and it reads plus fourteen point one percent for twenty twenty-five. Your verified figure is plus twelve point oh three. One minute: what do you do, and what happened? What you do: the verified figure wins pending investigation, and you read the generated DAX the way you read the inherited recipe in week two — what does each part actually do? What happened is almost certainly one of the bug classes you already know: an unequal window, or a wrong base — the Lesson Six audit toolkit names generated bugs exactly as well as human ones, because generated DAX fails in precisely the ways hand-written DAX fails. The deep point: renders-without-errors has never once meant correct in this course, and the arrival of AI does not amend that rule. It doubles down on it.

[Slide 5] Rapid tour of the rest. Forecasting extends your monthly line with a confidence band, under one central assumption: the future resembles the past's pattern. Our course data forecasts beautifully — and that is itself a teaching point, because it was generated with stable growth and clean seasonality, exactly the world the algorithm assumes. Real revenue meets pandemics and competitors the band knows nothing about. Ship forecasts with the band visible and the assumption stated, or you have built the truncated axis of the time dimension. Anomaly detection rings doorbells — and every flag gets the Lesson Three investigation before action, because May's file was anomalous too, until seasonality explained it. And Copilot: it drafts pages, which then face the full design pass; it writes DAX, which you read like inherited code and verify against a figure you know; and it summarizes visuals in fluent prose, in which every number gets checked — because a fabricated figure reads identically to a real one, and fluent has never meant faithful.

[Slide 6] Let me say the course's thesis one final time, because tonight is where it pays. Nothing in this course errors when it is wrong. Not the type change, not the inner join, not the truncated axis, not the forgotten filter — and not the generated answer. The common defense was never suspicion of tools. It was possession of expectations: a figure you know, a count you predicted, a claim you can state before you look. AI features raise the speed of producing answers, and they raise the value of your expectations by exactly the same amount. That is the trade, and you are equipped for it. Three roads from here: the PL-300 certification, which you are substantially prepared for; the Fabric platform, which is the engineering side of every star you consumed; and the third road, which was the real product all along — being the person in any room, against any confident number, from any source, who asks: average of what? Compared to which base? Excluding what? This week: the final practice trio ends with you writing the team's AI policy. Then the capstone. Build something true.