Meeting Summarisation

Human note-takers face a fundamental cognitive split: they must simultaneously listen, evaluate importance, and write. This divided attention means they capture explicit action items ('John will send the report by Friday') but miss implicit commitments ('I'll check those numbers' said in passing) and conditional agreements ('If the budget gets approved, we should start hiring').

There is also a recency bias. Note-takers capture more detail from the last twenty minutes of a meeting than the first twenty, because recent items feel more urgent and earlier items have faded from working memory. Decisions made early in the meeting — often the most strategic ones — get summarised in less detail than tactical items discussed at the end.

Social dynamics further distort note-taking. Commitments made by senior leaders are more likely to be captured than those made by junior team members, not because they are more important but because they command more attention. Side conversations, offhand remarks, and quiet agreements between two participants may be functionally invisible to the note-taker.

AI processes the entire transcript with uniform attention, extracting every statement that carries a commitment, decision, or action regardless of when it occurred, who said it, or how casually it was phrased. The result is a more complete and more equitable record that catches what human attention naturally filters out.

The most useful meeting summaries are not prose — they are structured records — a principle of effective output formatting. A table with columns for Decision Made, Owner, Deadline, and Dependencies is more actionable than a narrative paragraph. Each row becomes a trackable item that can be assigned in a project management tool, turning a meeting transcript into a work plan.

Separating 'decisions' from 'discussion' prevents a common failure mode where interesting conversations are treated as conclusions. A discussion about potentially changing the pricing model is not a decision to change it. Prompting the AI to distinguish between 'resolved decisions', 'open discussions requiring follow-up', and 'assigned action items' prevents this ambiguity from creating confusion downstream.

Including who said what for key decisions adds accountability. 'The team decided to delay the launch' is less useful than 'Sarah proposed delaying the launch by two weeks to address performance issues. Michael agreed contingent on updating the customer communication plan. Vote passed with no objections.' Attribution makes it possible to follow up with the right person and clarifies the conditions attached to each decision.

For recurring meetings — weekly standups, project reviews, sprint retrospectives — a consistent summary template ensures comparability over time. When every summary follows the same structure, it becomes possible to track patterns: which action items consistently carry over unfinished, which discussion topics recur without resolution, and where deadlines are regularly missed.

The simplest approach uses a transcription service (Otter.ai, Fireflies, or a platform's built-in transcription like Zoom or Teams) followed by pasting the transcript into Claude or ChatGPT with a structured prompt template. This works for teams that want summarisation without committing to a dedicated tool and allows full customisation of the summary format.

Dedicated meeting AI tools (Otter, Fireflies, Fellow, Granola) integrate transcription with summarisation and often connect to project management tools for direct action item creation. The trade-off is convenience versus control — dedicated tools handle the workflow automatically but offer less flexibility in how summaries are structured and what they emphasise.

For sensitive meetings where data privacy is a concern — see PII handling —, the transcript should be reviewed before AI processing. Remove or anonymise any information that should not leave the organisation — personnel discussions, financial details, strategic plans that are not yet public. The same data privacy principles that apply to any AI interaction apply to meeting transcripts, which often contain the most candid internal discussions.

Regardless of the tool, the most valuable practice is sending the AI-generated summary to all participants for review within 24 hours. This serves two purposes: participants can correct any misinterpretations (AI may misattribute a statement or misunderstand technical jargon), and the act of reviewing the summary reinforces accountability for the commitments it documents.

Transcription quality is the foundation, and it is not always reliable. Speaker identification errors (attributing a statement to the wrong person), mishearing technical terms, and poor audio quality all produce transcripts with errors that the summarisation model inherits. The summary will confidently attribute statements to the wrong people or misrepresent technical details if the underlying transcript is inaccurate.

Non-verbal communication is invisible to text-based summarisation. Sarcasm, hesitation, body language indicating disagreement, and consensus built through nodding rather than verbal agreement are all absent from the transcript. An AI might report unanimous agreement on a decision when the transcript shows no objections — but the room may have contained visible reluctance that would change the interpretation.

Context-dependent statements require human interpretation. When someone says 'we should definitely do that' in a tone dripping with sarcasm, the transcript reads as agreement. Similarly, inside references, abbreviated terminology specific to the team, and implicit references to ongoing projects may be misinterpreted by the AI. A human reviewer who was present at the meeting catches these nuances; the AI cannot.

For high-stakes meetings — board meetings, contract negotiations, disciplinary proceedings — AI summaries should be treated as draft notes that require human review and approval — applying the same safety-first principles as any AI-generated content, not as authoritative records. The speed and completeness benefits of AI summarisation are valuable, but the final record should carry a human's verification.

Individual meeting summaries are useful for immediate follow-up, but their cumulative value grows substantially when stored and indexed as a searchable knowledge base — a core knowledge management practice. A team that has six months of structured meeting summaries can query them for historical context: when was a particular decision made, who was involved, what alternatives were considered, and what constraints shaped the outcome. This institutional memory is otherwise locked in individual recollections that degrade over time.

Structured summaries are far more searchable than raw transcripts or prose notes. When every summary includes tagged decisions, named owners, and dated deadlines, a simple search can surface every commitment a particular team member has made, every discussion about a specific project, or every deadline set in a given quarter. This structure turns meeting archives from passive records into active decision-support tools.

Teams adopting this practice should establish a consistent storage location and naming convention from the start. Whether summaries live in a shared drive, a project management tool, or a dedicated knowledge base, the indexing system matters more than the storage medium. Consistent tagging of participants, projects, and decision categories makes the archive useful; inconsistent tagging renders it a pile of documents that nobody searches.