Turn Every Meeting into Verified Actions — Then Price the Meetings Themselves

Workflow area: Leadership Ops · Tier: Foundation · Hands-on time: ~25 minutes

An engineering leader's week is 15–25 hours of meetings, and the follow-through tax — writing the recap, chasing the owners, remembering what was actually agreed — is another 3–5 hours nobody budgets. AI collapses the recap to minutes. The failure mode is equally fast: transcripts garble names, models assign owners who were never asked, and a confidently wrong recap circulated to twelve people creates more cleanup than no recap at all. The verification pass is the playbook.

When to use — and when not to

• Use it for every recurring team meeting you runstaff meetings, planning sessions, cross-team syncs, anywhere decisions and actions are supposed to come out the other end.
• Use it for meetings you attend but don't runthe extraction tells you what you personally committed to before someone else's recap (or nobody's) defines it for you.
• Use the calendar audit quarterlyrecurring meetings accrete; nobody prices them; the audit is how you find the ₹40-lakh-a-year standing meeting hiding in plain sight.
• Skip it for 1:1s and people conversationsrecording a report's 1:1 changes what they will say to you; that channel runs on trust, not transcripts (see pb-02 for the right way to prep them).
• Skip it where consent is missing or recording is sensitivelegal reviews, security incidents, HR matters, and any call where a participant hasn't agreed to the bot. Announce the notetaker at the top of every call; if anyone objects, turn it off and type.

The recipe

1. 1.

   Enable the notetaker with consent built in. Turn on Teams Intelligent Recap, Zoom AI Companion, or Meet's "take notes for me" for meetings you own. Configure the joining announcement so every participant hears that recording is on, and say it out loud anyway at the top: "AI notes are on so I can stay in the conversation — flag anything you want kept out."

2. 2.

   Run the transcript through a fixed extraction prompt. Same template every time, so outputs are comparable week to week:

   Copy

   You are extracting a meeting record from a transcript. Meeting:
   [NAME, DATE, ATTENDEES]. My role: engineering leader who ran it.
   
   Extract, using ONLY what the transcript supports:
   
   1. DECISIONS — one line each, with who proposed it, who agreed,
      and the verbatim phrase that shows agreement. If agreement is
      ambiguous, mark [UNCONFIRMED] instead of guessing.
   2. ACTIONS — owner, action, deadline. Only name an owner if the
      transcript shows them accepting. Otherwise write "owner not
      assigned" — do not infer from who spoke most about the topic.
   3. OPEN QUESTIONS — raised but not resolved.
   4. PARKED — explicitly deferred items, with any revisit date.
   
   Rules: no section padding — write "none" where there is nothing.
   Flag anything you are unsure was said with [CHECK]. Do not
   summarize discussion; extract commitments only.
   
   Transcript:
   [PASTE TRANSCRIPT]
   

3. 3.

   Verify against your own memory before anything leaves your drafts. Two minutes, three questions: Did we actually decide that, or just discuss it? Did that person actually accept that action, or did the model assign it? Is any deadline invented? Good output looks like this:

   Copy

   DECISIONS
   - Ship the billing migration behind a flag on Aug 18 — proposed
     by Meera, agreed by Arjun ("yeah, flag it and go") [11:42]
   - [UNCONFIRMED] Retire the v1 API this quarter — discussed at
     length, no explicit agreement found.
   
   ACTIONS
   - Arjun — draft the flag rollout plan — by Aug 11
   - owner not assigned — customer comms for the migration
   

   The [UNCONFIRMED] and "owner not assigned" lines are the health signal. A recap where every decision is crisp and every action has an owner is usually a recap the model cleaned up into fiction.

4. 4.

   Circulate the verified recap and push actions into the tracker. Send the recap to attendees with a one-line correction window ("reply by EOD if I got anything wrong"), then create tracker tickets for each action — owner, deadline, link back to the recap. Actions that live only in a recap document die there.

5. 5.

   Once a quarter, run the calendar audit. Export your calendar (or your team's recurring-meeting list) and price it:

   Copy

   Here is a list of recurring meetings: name, cadence, duration,
   attendee count, and average loaded cost per attendee-hour of
   [RATE OR "use a placeholder of $X"].
   
   For each meeting compute annual attendee-hours and annual cost.
   Then recommend for each: KEEP (state the decision or coordination
   it uniquely produces), SHRINK (fewer people or minutes — say
   which), MERGE, MAKE ASYNC (what replaces it), or KILL. Rank by
   annual cost. Do not soften: if a meeting produces status that a
   dashboard already shows, say so.
   
   [PASTE MEETING LIST]
   

   Take the top three recommendations to the team as proposals, not decrees — the model prices meetings; it cannot see the trust or context a meeting quietly maintains. Kill one, shrink one, and re-run the audit next quarter.

Make it yours

• Small team or startup: skip the enterprise notetaker; record with consent, transcribe, and keep the whole pipeline in one Claude Project called "Meeting records" with the extraction prompt in the project instructions. The calendar audit matters more here, not less — a 10-person team in five recurring meetings is burning a headcount.
• Large org or GCC: use only the meeting AI inside your approved tenant (Teams Recap under your M365 agreement, not a consumer bot), confirm your retention policy covers transcripts, and standardize the extraction template across your leads so recaps are comparable in skip-level reviews.
• Distributed team across time zones: the verified recap becomes the meeting for whoever was asleep — add a "for async readers" line summarizing the one thing they must know, and hold decisions marked [UNCONFIRMED] open in the recap thread for 24 hours before treating them as made.

Quality check

• Every decision traces to a verbatim phrasespot-check two per recap against the transcript; agreement the model inferred is not agreement.
• No invented owners or deadlines"owner not assigned" appearing regularly means the extraction is honest; its absence means it is filling gaps.
• Corrections actually arriveif nobody has corrected a recap in a month, either you are superhuman or people stopped reading; ask.
• The calendar audit survives the team's sniff testverify the attendee counts and cadences with the people in the meetings before acting; a stale invite list inflates costs and torches your credibility.
• Actions land in the tracker, not just the recapthe pipeline's output is tickets with owners, and a recap with zero tracker links is theater.