Case Study 1: Dana's Memo — An AI Draft Rescued by Judgment

A worked before/after. Dana Whitfield, a data scientist, has run a churn analysis and must write a short memo to Renée Okafor, her company's VP of Marketing. She reaches for an AI to draft it — and discovers, the productive way, why the draft needs her. Fictional but realistic; the AI output is an illustrative composite of what current general-purpose models produce.

The temptation

It's late, the memo is due tomorrow, and Dana has the analysis done — she knows the finding cold. So she does the tempting thing: she opens an LLM and types the lazy prompt.

Prompt: "Write a memo to our VP of Marketing about our customer churn analysis."

Four seconds later she has this:

"Subject: Customer Churn Analysis — Key Insights

Dear VP of Marketing,

I am pleased to share the results of our comprehensive customer churn analysis. In today's competitive landscape, understanding customer retention is more critical than ever. Our in-depth analysis leveraged advanced analytics to uncover actionable insights that will empower data-driven decision-making.

Our findings reveal several key trends. Customer churn remains a significant challenge, and addressing it will be crucial for sustainable growth. We recommend a multi-faceted approach to improve retention and maximize customer lifetime value. By implementing targeted strategies, we can drive meaningful improvements in our key metrics.

Please let me know if you have any questions. I look forward to discussing these exciting findings."

It reads fine. It's grammatical, it's confident, and Dana almost pastes it into an email. Then she reads it as Renée would — and sees it for what it is.

What's actually wrong with it

The draft is fluent and empty — a textbook case of the §29.4 failure. Run it against what the book has taught:

• It contains no finding. "Several key trends." "A significant challenge." Which trends? How significant? The one thing a memo about an analysis must contain — the result — is exactly what's missing. The model couldn't supply it, because the model didn't do the analysis; it doesn't have Dana's data (the §29.3 institutional-knowledge gap, live).
• It's the generic default register throughout. "Comprehensive," "in-depth," "advanced analytics," "actionable insights," "empower data-driven decision-making," "exciting findings" — every phrase is the interchangeable AI voice, the bloat Chapter 3 taught Dana to cut. It could be a memo about any analysis at any company.
• It buries nothing, because there's nothing to bury. Chapter 27 (and Chapter 20 before it) drilled the discipline: lead with the recommendation, the "so what?" This memo leads with throat-clearing and ends with "let me know if you have questions" — because the model had no so what to lead with.
• It doesn't know Renée. Renée is impatient, reads on her phone, and wants the number and the ask in the first two lines. The model doesn't know that. It wrote for a generic "VP of Marketing," which is to say, for no one.

Dana realizes the draft isn't a starting point she can polish — it's a vacuum wearing a suit. So she does what the chapter teaches: she throws it out and starts over in the right order.

The rescue: think first, then let the model assist

Dana writes the substance herself — fast, because she knows it — in her own plain words:

"Rough: 60% of churn last quarter came from customers in their first 30 days. They're leaving before they ever hit our 'aha' feature (the dashboard). If we get first-week users into the dashboard, we should keep a lot of them. Recommend: a guided onboarding that pushes new users to the dashboard in week one. Pilot it, measure 30-day retention."

Now she brings the model in — not to write, but to revise, with a real brief:

Prompt: "Tighten this into a 120-word memo to our VP of Marketing, Renée. She's impatient and reads on her phone — lead with the recommendation and the 60% number in the first two sentences. Keep every specific (60%, first 30 days, the dashboard as the 'aha' feature, pilot + measure 30-day retention). Direct, no marketing fluff. Here's the voice she likes: 'Bottom line up top, then the why.' Give me two options."

The model returns options that keep Dana's substance and sharpen the prose. She picks one and edits the last line herself:

"Subject: Churn fix — target first-week onboarding

Renée — recommendation first: let's pilot a guided onboarding that pushes new users into the dashboard in their first week. Here's why. 60% of last quarter's churn came from customers in their first 30 days, and they're leaving before they ever reach the dashboard — our stickiest feature. Get them there early and we should retain a meaningful share of them. I'd run a four-week pilot on new signups and measure 30-day retention against the current cohort. If it moves, we scale it. Want me to scope the pilot?"

What changed, and what the model did (and didn't) do

The two memos are the same length and completely different documents. The first says nothing; the second leads with the ask, carries the real number, names the real mechanism, and proposes a concrete, measurable next step — the Chapter 27 "so what?" discipline, fully applied.

And here's the line that matters: the model never supplied a single piece of substance. The 60%, the first-30-days finding, the dashboard-as-aha insight, the pilot design — all Dana's, written before the model saw a word. The model did exactly what it's good at (§29.2): it tightened prose Dana had already filled with her own knowledge and pitched it to an audience Dana specified. The thinking happened where Chapter 1 says it must — in Dana's head, on her rough draft. The model made the typing faster, not the thinking absent.

That's the difference between the two prompts, and it's the whole chapter. The first prompt asked the model to think (it can't, so it produced fluent nothing). The second asked it to help express thinking Dana had already done (it can, so it produced a sharp memo). Same tool, same four seconds, opposite results — decided entirely by who did the thinking, and when.

The takeaway

Dana didn't avoid AI; she used it well, by refusing to let it do the one thing it can't. Her judgment supplied everything that made the memo good — the finding, the framing, the knowledge of Renée — and the model supplied speed on everything else. The draft that almost shipped would have wasted Renée's time and made Dana look like she had nothing to say. The draft that did ship was unmistakably Dana's, because the thinking was. Writing is thinking — and the tool that drafts can't do that part for you, no matter how fluent the first four seconds look.

Related: Chapter 29 · Case Study 2 · Exercises · Key Takeaways