Meeting Notes to Verified Action Items and Decisions

Tier: Foundation · Hands-on time: ~20 minutes

This playbook turns a raw transcript or messy notes into a verified list of both — where "verified" means every item is backed by a verbatim line from the transcript, so you never assign someone work they did not agree to.

When to use

• •
  You have a meeting transcript (from your conferencing tool's recording feature) or your own rough notes.
• •
  You are the one responsible for sending minutes, and people will act on what you send.
• •
  A long email or chat thread needs the same treatment — the workflow is identical.

Prerequisites

• •
  The transcript or notes as plain text. Export from your meeting tool, or paste your notes.
• •
  Consent norms: confirm your organization allows meeting recording/transcription, and that attendees knew.
• •
  Any current AI chat assistant. For sensitive internal meetings, use your enterprise-approved tenant or the local variant below.

The workflow

1. 1.

   Prepare the input. Paste the transcript. If it has speaker labels, keep them — they are what makes owner assignment reliable. If your notes lack speakers, add initials where you remember them.

2. 2.

   Extract with an evidence rule. Use this prompt:

   Copy

   From the meeting transcript below, extract:
   
   1. DECISIONS — things the group agreed to. Format: decision, who confirmed
      it, and the verbatim transcript line(s) showing agreement.
   2. ACTION ITEMS — table with columns: task, owner, due date, verbatim
      supporting quote. Owner and due date must come from the transcript;
      if not stated, write "UNASSIGNED" or "NO DATE" — do not guess.
   3. OPEN QUESTIONS — raised but not resolved.
   4. PARKED — explicitly deferred topics.
   
   Rules: every item must include a verbatim quote. If you are inferring
   rather than quoting, prefix the item with [INFERRED]. Do not merge
   similar-sounding items from different speakers.
   
   Transcript:
   <paste transcript>
   

3. 3.

   Audit the quotes. Search the transcript for 3–4 of the quoted lines, always including: every decision, and any action item with a named owner and date. Verbatim match required.

4. 4.

   Resolve the gaps yourself. Items marked UNASSIGNED, NO DATE, or [INFERRED] are your to-do list as the note-taker: confirm them with the relevant person before publishing, or list them explicitly as "needs owner."

5. 5.

   Format for sending. Ask:

   Copy

   Rewrite the verified list as concise meeting minutes: Decisions,
   Action Items (owner - task - due date), Open Questions. Drop the quotes.
   Neutral professional tone, no filler.
   

   Read it once end-to-end before sending — you are the author of record.

Tool variants

• Meeting platform's built-in AI recap: fastest, but recaps typically skip the evidence step — paste the platform's transcript into a chat assistant and run steps 2–3 anyway when accuracy matters.
• General chat assistant (any major vendor): the full workflow as written; best quote fidelity.
• Local model via Ollama: works well for this task even at 8B scale because extraction needs no world knowledge; split transcripts over ~1 hour into halves.

Verification

• •
  Every decision's supporting quote found verbatim in the transcript.
• •
  Every named owner actually spoke (or was addressed) in the quoted line — models sometimes attribute a task to whoever mentioned the topic.
• •
  Count check: skim the transcript for the words "I'll", "I will", "by Friday", "agreed", "let's go with" — each hit should map to an extracted item or be consciously excluded.
• •
  No [INFERRED] items remain in the published minutes without human confirmation.

Pitfalls

• Attribution drift. "Priya asked whether the migration could ship Friday" becoming "Priya will ship the migration Friday" is the classic failure — the evidence-quote rule catches it, but only if you audit the quote.
• Transcription errors upstream. Auto-transcripts mangle names and numbers ("Anvi" → "envy", "$15K" → "50K"). Verify numbers and names against your own memory, not the transcript alone.
• Summarizing instead of extracting. If the output reads like prose, the model ignored your structure — restate "output only the four sections, no narrative."
• Publishing without reading. Minutes carry your name. One misassigned task erodes trust in every future AI-assisted summary you send.

Restricted-environment variant

For client calls or meetings covered by NDA:

• •
  Replace client and person names with role labels (CLIENT_PM, OUR_TL) before pasting into a hosted tool; keep the mapping locally and re-substitute after.
• •
  Strip commercial figures and anything legally sensitive; extract those lines manually.
• •
  Full-privacy path: run the identical prompts on a local Ollama model (pb-07) — transcript extraction is one of the tasks where local 8B models are genuinely sufficient.
• •
  Check whether your client contract prohibits sending call content to third-party processors — many India services MSAs do.

Success criteria

• •
  Minutes sent within 30 minutes of the meeting ending.
• •
  Zero corrections from attendees on decisions or ownership ("that's not what I agreed to" never appears in replies).
• •
  Every action item has an owner and a date, or is explicitly flagged as needing one.
• •
  You spent your review time on the flagged gaps, not on re-reading the whole transcript.