Run three meeting transcripts through an AI pipeline — decision log, owner-dated action table, tracker entries, weekly status draft — then audit every line. One transcript hides a reversed decision, one action has no owner the AI will confidently invent, and one sentence must never leave the room. Catching all three is the graded outcome.
Meera Nair is Director of Engineering at Corvid Systems, a fictional 45-engineer logistics SaaS company. Her Mondays produce three meetings and zero written follow-through: decisions live in people's heads, actions evaporate, and her weekly status email goes out Thursday, late and vague. This lab builds her — and your — leadership-ops pipeline: transcript in, circulated-ready status draft out, with a human verification gate at every stage. The pipeline is only trustworthy if you can prove it catches what AI summarizers reliably get wrong: recording the loudest voice instead of the final decision, assigning ownerless work to whoever mentioned it, and leaking sensitive people-context into documents that circulate.
Three condensed excerpts. Paste-ready.
TRANSCRIPT 1 | Eng staff meeting, Monday 09:30 (Meera + 5 leads)
Arjun (Platform lead): "We should move to weekly releases. Biweekly is
killing us. I've said this for two months — weekly, starting now."
Divya (QA lead): "Regression suite takes 9 hours. Weekly is possible
only if we cut the suite or parallelize it."
Arjun: "Then parallelize it. Weekly releases. Everyone agrees, right?"
Tomás (Mobile lead): "App store review makes weekly pointless for us."
Meera: "Hold on. The SOC 2 audit window opens October 6. Change-control
evidence gets sampled from release records. I'm making the call: we
STAY on biweekly until the audit closes in November, then revisit
weekly with a parallelized suite as the precondition. Divya, scope the
parallelization by Sept 12."
Arjun: "Fine. But I want it on record I think weekly is right."
Meera: "Noted. Also — someone should update the on-call handbook before
the new joiners start rotation. It still references PagerDuty, we're
on Opsgenie now."
Priya (Infra lead): "It's badly out of date. I can't take it though —
I'm heads-down on the VPC migration until the 20th."
Meera: "Understood. We'll find an owner. Last thing, and this stays in
this room: Sam on Tomás's team is dealing with a family medical
situation. Keep his load light through September. Not for any doc."
TRANSCRIPT 2 | Incident review, Monday 13:00 (INC-4412, checkout outage)
Facilitator: "41 minutes of checkout degradation Saturday. Root cause?"
Priya: "Connection pool exhaustion after the payments deploy. Rollback
fixed it in 6 minutes once we decided — deciding took 28."
Divya: "The runbook said page the DBA. We don't have a DBA anymore."
Actions agreed: Priya to add pool-saturation alert by Sept 10. Divya
to purge stale runbook steps by Sept 17. Tomás to add checkout
synthetic check by Sept 12. Retro doc to be shared org-wide.
TRANSCRIPT 3 | Cross-team planning, Monday 15:30 (with Product)
Nadia (PM): "Freight-quote API for the Kestrel deal — can we commit to
October 15?"
Arjun: "Platform can deliver the API surface by Oct 8 IF the rate-table
service is stable. That's a dependency on Priya's migration."
Priya: "Migration lands Sept 20. Tight but yes."
Nadia: "So committing Oct 15 externally?"
Meera: "Commit Oct 15 to Kestrel with a contractual grace window to
Oct 22. Arjun owns the API delivery. Nadia drafts the client comms
by Friday."
Read all three transcripts yourself first — ten minutes, no AI. Write down, in your own words: every decision you believe was actually made, every action with its owner, and anything that should not appear in a circulated document. This pre-AI pass is your anchor; the audit steps only work if you formed a view first.
Generate the decision log. Paste all three transcripts with this prompt: "You are a chief of staff. Extract a decision log from these three meetings. For each decision: the decision as finally made, who made the call, the date it applies from, and any conditions attached. Decisions only — not proposals, not opinions." Save the output unedited.
Audit the decision log against the transcripts. For each logged decision, find the transcript lines that support it. Check one thing with maximum suspicion: where discussion went back and forth, did the log record the final ruling or the most-repeated proposal? Mark each decision VERIFIED or WRONG with the supporting line. If the release-cadence entry says anything other than a hold, you have found the first trap.
Generate the owner-dated action table. Prompt: "Extract every action item as a table: Action | Owner | Due date | Source meeting. If the transcript does not name an owner or date, write UNASSIGNED or NO DATE — do not guess." The instruction not to guess is deliberate; you are testing whether the model obeys it.
Audit every owner and date. For each row, confirm the transcript explicitly names that person accepting that work. Watch the on-call handbook row: Priya described the problem and explicitly declined ownership. If the table lists Priya — or anyone — as owner, mark it FABRICATED, correct it to UNASSIGNED, and add a real next step: "Meera to assign at Thursday staff meeting."
Draft tracker entries. Ask the AI to convert only VERIFIED actions into tracker-ready entries: "For each verified action, write a ticket title, a two-line description, owner, and due date." Ownerless items get a triage ticket assigned to the meeting chair, not a delivery ticket assigned to a guessed name.
Draft the weekly status email. Prompt: "Draft a weekly engineering status email for company-wide circulation from this decision log and action table. Sections: Decisions, Delivery commitments, Incident follow-ups, Risks. Crisp, factual, no praise language." Note the audience: company-wide.
Run the privacy scrub. Reread the draft line by line against one rule: nothing said in confidence appears in circulated output. Transcript 1 contains a sentence Meera explicitly fenced off. If any trace of Sam's situation — even a euphemism like "temporary capacity reduction on mobile" attributed to a person — appears in the status draft or tracker entries, remove it entirely. Capacity planning may be discussed with Tomás privately; it does not go in the email.
Finalize and self-grade. Assemble the four artifacts: verified decision log, corrected action table, tracker entries, scrubbed status draft. Write a three-line audit note on top: what the AI got wrong, what you corrected, what you removed. That note is the habit this lab exists to build.
Summarization models weight repetition as importance — a proposal argued three times reads as more decided than a ruling stated once. Always audit contested decisions against who had authority and what was said last.
A complete lab has all of the following:
An AI meeting pipeline is a drafting engine, not a system of record. The three gates — final-ruling check, owner-verification check, privacy scrub — are what turn fast summaries into documents a leader can circulate without a correction email the next day.