Three workshop transcripts, one AI-drafted findings pack — and seven defects hiding in it: five citations pointing to the wrong speaker and two "needs" nobody ever stated. Catch all seven with a line-by-line audit, then rebuild a pack you could defend in front of the people who were in the room.
Meera Pillai is a senior business analyst at TransHarbor Logistics, a fictional freight forwarder headquartered in Chennai with a European hub in Rotterdam. The company is replacing its email-and-spreadsheet vendor onboarding process with a portal, and Meera ran three elicitation workshops last week: one with the sponsor, one with operations, one with IT. A colleague already fed the transcripts to an AI and circulated a findings pack. It reads beautifully. Meera's instinct — and the discipline this lab drills — is that a findings pack is worthless until every requirement candidate has been traced back to a speaker and a line.
Three condensed excerpts with line numbers. Paste-ready.
W1 | Sponsor workshop — Rajan Iyer, VP Commercial
W1-01 "This portal has to be live before the Hanover trade fair in March."
W1-02 "Vendors should self-register, upload their compliance docs, and
W1-03 track their own approval status. No more email ping-pong."
W1-04 "While we are at it, I want vendor performance dashboards, a rate-
W1-05 comparison module, and maybe supplier auctions in phase one."
W1-06 "Our Chinese carrier partners keep asking for Mandarin support —
W1-07 add that to the list."
W2 | Operations workshop — Elke Brandt, Ops Manager, Rotterdam
W2-01 "Rajan keeps saying we onboard forty vendors a month. It is forty
W2-02 a QUARTER. Sixty in peak quarter, and I have the tracker to prove it."
W2-03 "The real pain is compliance-document chasing — insurance certs
W2-04 expire and nobody notices until a shipment is blocked."
W2-05 "My team approves vendors from the warehouse floor half the time.
W2-06 Whatever you build must work on a phone browser."
W2-07 "Do not give me auctions. We tried a tool in 2023; carriers hated it."
W3 | IT workshop — Sanjay Rao, Enterprise Architect
W3-01 "Vendor master data lives in SAP. The sync interface is batch —
W3-02 it runs nightly at 02:00 and that will not change this year."
W3-03 "Single sign-on is Azure AD only. No local accounts, non-negotiable."
W3-04 "There is no integration budget left this FY, so whatever you scope,
W3-05 it talks to what already exists."
W3-06 "Nobody has answered my question about who owns vendor data quality
W3-07 once vendors can edit their own records."
This is the AI-drafted pack your colleague sent around. It contains the defects you are hunting. Do not correct it yet — audit it first.
FINDINGS PACK v0.1 — Vendor Onboarding Portal (AI-drafted)
Theme A — Self-service onboarding
RC-01 Vendors self-register and upload compliance docs (Rajan, W1-02)
RC-02 Vendors need a native mobile app for onboarding (Elke, W2-05)
RC-03 Portal must support Mandarin (Elke, W2-06)
Theme B — Volumes and timing
RC-04 Go-live before Hanover fair in March (Rajan, W1-01)
RC-05 We onboard ~40 vendors per month (Rajan, W1-01)
Theme C — Compliance management
RC-06 Auto-alerts before insurance certificates expire (Elke, W2-03)
RC-07 Automated vendor credit scoring at registration (Sanjay, W3-04)
Theme D — Technical constraints
RC-08 SAP sync is real-time via API (Sanjay, W3-01)
RC-09 SSO via Azure AD only, no local accounts (Rajan, W1-03)
RC-10 Data-quality ownership is unresolved (Sanjay, W3-06)
Read the transcripts yourself first — ten minutes, no AI. Write down, in your own words, what each speaker actually asked for, and note the one open conflict between two speakers (there is at least one — volumes). This pre-AI anchor is what lets your instinct fire later when a citation smells wrong.
Build the audit log. Create a table with columns: RC # | Claim | Cited speaker+line | What that line actually says | Verdict | Notes. Verdict values: CLEAN, MISATTRIBUTED (real statement, wrong speaker or wrong line), INVENTED (no speaker said it anywhere). This log — not the findings pack — is the artifact this lab grades.
Audit RC-01 through RC-10 by hand. For each candidate, go to the cited line and read it. Ask two questions separately: does this line say it, and does this speaker say it anywhere? A claim that is true but cited to the wrong person is still a defect — in a real review, Rajan will be asked to defend statements he never made. Expect roughly half the rows to fail.
Cross-examine with AI. Paste the transcripts and the findings pack with this prompt: "Audit this findings pack against these transcripts. For each RC, quote the exact transcript line that supports it, or state NO SUPPORT. Flag any RC where the cited speaker never said it." Compare the AI's audit to yours. Where you disagree, the transcript is the referee — reread the line. The AI will usually catch the inventions but be lazier about wrong-speaker misattributions; that gap is the lesson.
Resolve the volume conflict explicitly. RC-05 cites Rajan for a monthly figure. What does W2-01/02 say, and who has evidence? Write the conflict into your log as its own row: both quotes, both line numbers, and a follow-up action ("get Elke's tracker") — do not silently pick a winner.
Rebuild the findings pack. Prompt: "Regenerate this findings pack using only claims I have marked CLEAN, plus the corrected versions I list below. Every requirement candidate must cite speaker and line. Add a section called 'Open conflicts and unanswered questions' for the volume dispute and Sanjay's data-ownership question." Supply your corrections (right speaker, right line, inventions deleted).
Pressure-test the rebuilt pack once. Ask the AI: "Is there any candidate in this pack that a speaker in the transcripts explicitly argued against?" Check its answer against W2-07 — if auctions or anything phase-one-maximal survived into your pack without a flag, fix it.
Write the three-line audit summary. At the top of the log: how many candidates audited, how many misattributed, how many invented, one conflict logged. This is the sentence you would say before presenting the pack: it is what makes the room trust the rest.
Audit the speaker and the line as two separate checks. "Somebody did say this" is exactly how misattributions survive review — the claim feels familiar, so nobody checks whose mouth it came from.
A complete lab has all of the following:
A findings pack earns trust through its citations, not its prose. Audit speaker and line as separate checks, log conflicts instead of resolving them by vibe, and delete anything no human actually said — the AI will happily invent the requirement the pattern suggests.