Practice — Hot · Solution

A policy that would hold (one strong version)

Analytics AI Usage Policy — v1

Generated DAX ships only after (a) the reviewer reads the method — every function explained aloud — and (b) the result matches a manifest figure or an independently computed check. (Evidence: the 0.42 cost ratio and the index-not-growth formula both rendered plausibly.) Generated pages pass the full design review — question test first — like any junior analyst's draft. Generated prose: every number and every claim verified independently against the visuals; truth in one sentence licenses nothing about the next (evidence: "12.0% growth, led by the West's 112.7% growth"). Q&A ships to stakeholders only with tested synonyms, and any answer quoted onward requires reading the field readout (evidence: "average revenue," which is two numbers 2× apart). Influencer findings are written as associations with base sizes stated; the word "drives" requires an experiment we did not run. Forecasts travel with band, assumption sentence, and visible history boundary — no exceptions for slide aesthetics.

Escalation: when generated output disagrees with a verified figure, the verified figure wins pending investigation; the disagreement is logged and the generated method is read until the discrepancy is named. Never average, never pick the flattering one.

Lean in: first drafts of pages and measures (reviewed as above — the draft is free, the review was always the job), and stakeholder Q&A on configured models, where our naming work makes the parser trustworthy. The tools are fast interns: hire them, review them, sign what ships. — [Analyst], for the analytics team

Reflection answers

• The match-a-verified-figure rule catches the most: artifacts 1, 6, and the +14.1% YoY all die on contact with a known number, and artifact 2's false half dies on checking each claim independently. One habit, most of the protection — which is why it led the policy.
• Because the course's product was never formulas — those are in the reference pages. The product is the discipline: expectations held against confident numbers, from any source. A policy memo is that discipline made transferable to people who never took the course — which is what working analysts actually do with what they know. The last exercise is the first day of the job.