Back to course overview
Module 2Analyzing data with AI 12 min

From question to evidence

The findings memo: every claim wired to a number, every number wired to a source, uncertainty stated like a professional — the artifact that survives hostile review.

Analysis that lives in your head, a chat log, and six spreadsheet tabs is not yet work product. The bridge between analysis and influence is the findings memo — one or two pages with a specific internal contract: every claim carries a number, every number carries a source, and confidence is stated explicitly. This is the document your recommendation stands on when someone senior pushes back — and someone senior always pushes back.

The structure (steal it)

  1. The question and the definition. One sentence each. 'Why did refund cost rise ~40% QoQ? Refund cost = refunded order value as % of completed order value, by refund month (definition approved by COO, June 30).'
  2. The finding. The cause, localized: which segment, since when, how big. Numbers inline, not 'significantly' or 'sharply' — 'Gear web refunds rose from ~10% to ~27% of Gear web revenue between March and June; all other category-channel cells moved less than 1.5 points.'
  3. The evidence chain. The 3-4 cuts that support it, each one line + source: metric by category (source: June orders extract, tab 2), reason mix in the suspect segment, the trend-break date, the counter-segment check. A reader should be able to re-derive your finding from this section alone.
  4. What it isn't. The alternative explanations you tested and killed, each with the cut that killed it. This section converts skeptics faster than the finding does — it shows you attacked your own answer before they could.
  5. Confidence and gaps. What you're sure of, what's probable, what you couldn't check ('fragile-flag audit pending — catalog access requested'). Stated limits raise credibility; discovered limits destroy it.
Prompt to try

Here is my findings memo draft. Review it as a hostile CFO: (1) flag every claim that lacks a number or a source; (2) flag every place I've stated an inference as a fact; (3) list the three hardest questions you'd ask in the meeting, and check whether the memo already answers them. Do not soften anything - I need the attack, not encouragement.

AI as red-team reviewer is the highest-value use in this whole module. It will find the unsourced claim you stopped seeing four drafts ago.

The chain of custody habit

Keep one running 'numbers log' per investigation: every statistic that might be quoted, its exact value, the file/tab/filter it came from, and the date pulled. Thirty seconds per entry. When the COO asks 'where does the 19% come from?' three weeks later, you answer in one breath — and that answer, delivered instantly, is worth more to your reputation than the analysis itself. (Teams with a governed metrics layer get this for free — certified definitions with owners. Until then, the log is your layer.)