Back to course overview
Module 1AI for discovery & requirements 18 min
Lab: The discovery pack
Run the full discovery workflow on three Harbor Lane interviews: grounded synthesis, cross-interview patterns, verification checklist, and the investigation requirements doc.
Your first week on the refund investigation: three interviews are done, and your job is to turn them into a discovery pack — themes with receipts, a data-verification checklist, and requirements for the analysis itself. Below are the transcript excerpts (fictional, but shaped like the real thing: partial memories, confident guesses, and one detail that will matter a lot).
interview 1 — COO (sponsor), 12 min excerpttext
COO: Finance flagged it in the quarterly review. Refunds used to run maybe
two, two-and-a-half percent of revenue. Last quarter it was past four.
MAYA: When did you first hear about it?
COO: The June close. But honestly it could have been building - nobody
watches that line monthly. I want to know if it's quality, fraud, or just
growth doing this.
MAYA: What would a good outcome look like?
COO: A cause I believe, a dollar figure, and something we can do about it
that doesn't torch customer experience. Returns are part of why people
trust us. I don't want a crackdown - I want to stop paying for whatever
is actually broken.interview 2 — support team lead, 15 min excerpttext
LEAD: The tickets changed in the spring. More 'arrived broken', 'box was
crushed', that kind of thing. My agents started a slack emoji for it.
MAYA: Broken items - any pattern in what products?
LEAD: The merch side, I think. Mugs, the pour-over kits, kettles. You
don't get 'arrived broken' on a bag of beans.
MAYA: Online orders, store purchases, or both?
LEAD: Shipping damage is online by definition. Store returns look normal
to me - it's the web tickets that moved.
MAYA: When in spring, if you had to pin it?
LEAD: April? May? I could pull ticket tags if you want actual numbers -
my 'spring' is a vibe, not a date.interview 3 — fulfillment manager, 10 min excerpttext
MGR: Nothing changed in our process. Same team, same pick-pack flow,
same carrier contracts since January.
MAYA: Any changes in materials, packaging, anything physical?
MGR: Well - procurement switched us to the lighter eco mailers in April.
Corporate initiative, cut packaging spend about thirty percent. But
that's not a process change, that's just boxes.
MAYA: Do the mailers handle everything? Fragile items too?
MGR: Everything except what's flagged fragile in the system. Although -
honestly I'd have to check what's actually flagged. The flags are old.
MAYA: Who would know?
MGR: That'd be whoever set up the product catalog. Years ago. Half the
gear items came in after that.- 1Synthesize each interview with the quote-grounded prompt from the 'Interviews & synthesis' lesson. Check every quote the AI returns against the transcript — Ctrl-F (Cmd-F) each quoted phrase in the source transcript; if it isn't a literal string match, the AI paraphrased inside quote marks — reject it, tighten the prompt, and rerun. (This check takes two minutes and builds the habit that saves you in front of a VP.)
- 2Run the cross-interview pass: agreements, contradictions, and checkable claims. You should surface: support's 'spring' vs. fulfillment's 'nothing changed' vs. the April mailer switch; the fragile-flag uncertainty; and the COO's 2.5%→4% figures as claims needing data.
- 3Build the verification checklist — every factual claim with the data that would confirm it (refund rate by month; refund reasons over time; 'arrived broken' by product category; web vs. store split; which SKUs are flagged fragile). This checklist is Module 2's work order.
- 4Draft the requirements doc for the investigation: the signed metric definition (write one and note you'd get the COO to approve it), must-answer questions, and out-of-scope items (e.g., fraud analysis unless data points there). Run the ambiguity-hunt pass on your own doc before calling it done.
- 5Form — but label — your hypothesis. A good analyst leaves week one with 'lighter mailers + stale fragile flags → transit damage on gear items' as a hypothesis, clearly marked, with the data that would confirm or kill it. Writing it down unlabeled is how confirmation bias gets a head start.
Problem set 1
You get two more transcripts (a store manager who contradicts the support lead, and a procurement analyst with an incentive to defend the mailer decision) plus a stakeholder map exercise: who has information, who has incentives, and whose 'facts' need the gentlest independent checking. Interviewing is easy; interviewing people with stakes is the job.