AI for meetings
From capture to useful notes
Useful notes compress a meeting without collapsing facts, interpretations, proposals, and decisions into one category.
Before you start
Why this matters
A transcript of a forty-five-minute meeting may contain eight thousand words. Sending it to absent colleagues transfers the work of interpretation to them. Asking AI to “summarize this” can create the opposite problem: five polished bullets that omit disagreement, turn suggestions into commitments, and hide uncertainty. Good notes sit between those extremes. They preserve what readers need to understand and act.
The goal is not maximum compression. It is faithful, useful structure.
1Learn the idea
Read
Define the note contract
Before generation, specify the output sections and the standard for each. A practical note contract includes:
- Purpose and context: why the meeting occurred and what material governed it;
- Key points: important facts and arguments, attributed where relevant;
- Decisions: choices explicitly made by authorized participants;
- Actions: tasks with an owner and a due date, or a visible missing field;
- Open questions: unresolved items that affect progress;
- Risks and disagreements: material concerns, including minority views;
- Sources: transcript timestamps, documents, or meeting artifacts used.
Different meetings need different contracts. A brainstorming session should preserve options without pretending they were selected. An incident review needs a careful timeline and confidence labels. A one-to-one may require minimal notes and stricter access. A board meeting may have formal minute requirements that an AI draft cannot satisfy without authorized review.
Read
Separate extraction from writing
One-pass summarization makes errors hard to detect. Use two stages.
First, ask the assistant to extract candidate items into a structured table or list. Each candidate should contain the statement, category, speaker if relevant, timestamp or source, and uncertainty. Do not ask for elegant prose yet.
Second, review or programmatically check those candidates, then ask the assistant to draft the reader-facing notes from the approved items. This approach exposes whether a “decision” has evidence and whether an action has an owner before fluent writing hides the gap.
A useful extraction instruction is:
Extract candidate facts, proposals, decisions, actions, questions, and risks. For each item, quote or closely reference the supporting passage and timestamp. Do not infer agreement from silence or positive language. Mark missing owners, dates, and unclear speakers. Do not add information from general knowledge.
The transcript remains evidence, not absolute truth. If the capture was wrong, a timestamp only lets a reviewer locate the error faster.
Read
Choose the right level of detail
Write for a named reader and use case. An attendee may need a brief recap and action list. An absent executive may need the decision, rationale, tradeoffs, and major risks. A project team may need exact dependencies and dates.
Layering works well:
- a three-to-five-line executive summary;
- verified decisions and actions;
- discussion themes and unresolved issues;
- source links or timestamps for verification.
This lets readers scan without deleting the path back to evidence. Avoid chronological play-by-play unless sequence matters. Grouping discussion by topic usually makes the notes easier to use, but retain chronology for incidents, negotiations, or any meeting where one event changed the meaning of another.
Read
Preserve attribution carefully
Attribute statements when identity matters: a decision owner accepted a task, an expert raised a safety concern, or participants disagreed. Do not attach every ordinary observation to a person; that creates clutter and unnecessary personal data.
Never infer a speaker’s motive. “Marta opposed the plan because she fears change” is an interpretation unless Marta said exactly that. Prefer “Marta opposed the 1 August date, citing incomplete security testing.” The second version records a position and stated reason.
Use neutral language. Replace “The team finally admitted the forecast was unrealistic” with “The team revised the forecast from 400 to 280 accounts after reviewing current conversion data.” Notes should not silently amplify the prompt writer’s bias.
Read
Handle contradiction and uncertainty
Meetings often contain conflicting numbers or accounts. A weak summary chooses one. A strong summary records the conflict:
Revenue estimate remains unresolved: Finance cited £1.2 million; Sales cited £1.5 million. Omar will reconcile definitions by 7 May.
If the group did not resolve a disagreement, do not manufacture consensus. Phrases such as “the team discussed,” “two options remain,” and “no decision was recorded” are valuable. A summary can be useful precisely because it reveals what is incomplete.
Likewise, absence is not agreement. A participant who stayed silent did not necessarily approve. Casual acknowledgements—“sounds good,” “fine for now,” or a thumbs-up reaction—may need confirmation before being recorded as a consequential decision.
Read
Edit the AI draft
Review generated notes in a risk-first order:
- decisions and commitments;
- owners, dates, amounts, and names;
- disagreement, caveats, and open questions;
- context and rationale;
- style and brevity.
Do not spend ten minutes polishing headings before checking whether the launch date is wrong. Compare high-impact items to source passages. Ask an appropriate attendee to confirm decisions when the record is ambiguous.
Remove unsupported filler such as “The team had a productive discussion” unless it conveys operational meaning. Delete repeated points. Expand compressed wording when it changes nuance. Make every correction in the notes, not only in a reply that future readers may miss.
Read
Publish with status and provenance
Label the document “AI-assisted draft,” “reviewed notes,” or “approved minutes” according to its actual state. Include meeting date, note owner, last review time, and source location where policy allows. Control access before sharing.
Set a correction window: for example, “Please report factual corrections by 3 p.m. Wednesday; action owners should respond sooner if an assignment is wrong.” A deadline prevents indefinite ambiguity while preserving a route to repair.