Six months in, nobody remembers why the team chose the API integration over the file drop — but a meeting decided it, with reasons. This playbook pipes meeting transcripts through a fixed extraction into decisions-with-rationale, owner-dated actions, and RAID updates, audited by you before distribution — accruing the institutional memory projects usually lose.
Workflow area: Stakeholder Communication · Tier: Foundation · Hands-on time: ~30 minutes
Meetings are where projects actually decide things, and meeting outcomes are the worst-preserved artifact in business analysis. Minutes, when written at all, record attendance and agenda, not the rationale that makes a decision defensible a quarter later. Meeting AI plus a disciplined extraction template changes the economics: every workshop, steering session, and standup ritual can yield a structured record — what was decided and why, who owns what by when, what risks moved — in five minutes of analyst review instead of forty of typing. The audit pass matters because extraction models resolve ambiguity optimistically: a "let's probably go with B" becomes "Decision: B" unless a human who was in the room says otherwise.
Set up capture once. Enable your meeting AI (Teams Copilot, Zoom AI, Fireflies) for the meetings you own, with the consent announcement on. Decide where logs live — a Confluence space, Notion database, or SharePoint list with one entry per meeting and running registers for decisions, actions, and RAID.
Run the transcript through the fixed extraction.
You are extracting a structured record from a meeting transcript.
Meeting: [NAME, DATE, ATTENDEES + ROLES]
Project context: [ONE LINE]
Prior open actions: [PASTE CURRENT ACTION REGISTER, or "none"]
Extract, using ONLY the transcript:
1. DECISIONS — for each: the decision in one sentence; the
RATIONALE as stated (quote or close paraphrase with speaker
and timestamp); alternatives that were discussed and set
aside; who made or ratified the call. If something was
discussed but NOT clearly decided, list it under DISCUSSED,
NOT DECIDED — do not promote leanings into decisions.
2. ACTIONS — owner, action, due date. If owner or date was not
stated, write [UNASSIGNED] / [NO DATE] — never guess.
3. RAID UPDATES — new or changed risks, assumptions, issues,
dependencies, each with the timestamp where it surfaced.
4. PRIOR ACTIONS — status of each open action from the register,
if mentioned: done / progressing / stuck / not mentioned.
5. PARKING LOT — raised but deferred topics.
Rules: every item carries a timestamp. Never merge two speakers'
statements. Mark anything you are uncertain about [CHECK].
Read a sample of good output.
DECISIONS
- D-14: Vendor data flows via API integration, not nightly file
drop. Rationale: "the file drop means duplicates live for 24
hours and month-end lands on a weekend" (Ops lead, 22:41).
Alternatives set aside: file drop (latency), manual entry
(volume). Ratified by: program sponsor (24:05).
DISCUSSED, NOT DECIDED
- Retry behavior on API failure — two options aired, no call
made (31:10). → parking lot.
ACTIONS
- A-31: IT architect — confirm API rate limits with vendor —
due 21 May.
- A-32: [UNASSIGNED] — draft rollback comms — [NO DATE] [CHECK]
Audit before distribution — the five-minute pass that makes it trustworthy. You were in the meeting; the model was not. Check three things: every DECISION was actually decided (the most common error is promoting a strong opinion into a ruling); every rationale quote is faithful at its timestamp; every [UNASSIGNED] and [CHECK] is resolved — by you assigning from knowledge of the room, or by an explicit follow-up ask. Then distribute to attendees with a 48-hour correction window: "this becomes the record on Thursday."
Merge into the running registers. Decisions append to the decision log with IDs; actions merge into the action register; RAID updates patch the RAID log. The IDs matter — requirements can now cite D-14 as their origin (feeding pb-04's evidence trails), and the CCB can ask "what did we decide about retries?" and get an answer.
Exploit the accumulated memory. Once ten meetings are logged, the registers become a queryable corpus: paste the decision log into Claude and ask "list every decision touching the vendor API, with rationale" before the next architecture session. Onboarding a new analyst now starts with "read the decision log," not "ask around" — this corpus is what pb-17's grounded Q&A is built on.
Meeting transcripts capture candor — disagreements, names, half-formed positions — that the polished record should not; process transcripts only in your organization's approved AI tenant, distribute the audited extraction rather than raw transcripts, and apply your retention policy to the recordings themselves.
A decision without its recorded rationale will be re-litigated; five minutes of audited extraction per meeting buys a program memory that answers "why did we choose this?" for as long as the project lives.