What Meeting Summaries Get Wrong

Meeting transcription and summarisation now works well enough to be default in most workplace tools. The quality is good and there is one systematic error worth knowing about.

The action item problem

Attribution is the weak point. Automatic action items are frequently assigned to whoever spoke about a topic most, or last, rather than whoever agreed to own it.

The pattern is predictable: a manager describes a problem at length, someone else says "I'll pick that up", and the action lands on the manager. Or a discussion where three people contribute and the task is assigned to the one who summarised.

Check the action list before it is circulated. This is thirty seconds and it prevents the most common workplace consequence of these tools: work that nobody does because everyone thought it was assigned elsewhere.

Other systematic errors

Decisions recorded that were not made. Discussion of an option gets summarised as agreement on it. Where a meeting explored and did not conclude, summaries tend to manufacture a conclusion.

Confidence flattening. "We could maybe try X" becomes "the team will try X". Hedged discussion is compressed into definite statements because summaries are written definitively.

Quiet participants disappear. Someone who spoke once, briefly, and made the key point may not appear at all. Volume of speech drives inclusion.

Disagreement smoothed over. A meeting with real conflict is summarised as a discussion that reached alignment. This is the most misleading failure, because the disagreement was the important content.

Names and jargon. Transcription errors on proper nouns and technical terms, carried into the summary.

Using them well

Keep the transcript, not just the summary. The summary is a convenience. When something is disputed later, you need the record.

Correct the action items immediately, in the meeting or right after, while people remember.

Confirm decisions explicitly during the meeting. "So we have decided X, and Y owns it" produces a clean line in the transcript and removes the ambiguity the summariser would otherwise resolve wrongly.

Circulate to participants and invite correction rather than treating the output as the record.

Do not use summaries for anything contested. Performance discussions, disputes, negotiations. Take real notes.

The consent question

Recording other people is regulated, and the rules vary substantially by jurisdiction. Some places require only one party's consent, others require everyone's. For a call spanning several countries, assume the strictest rule applies.

Announce it. Beyond the legal position, a tool that joins meetings silently and transcribes is a trust problem that surfaces badly.

Give people a way to decline. Someone who wants to raise something sensitive needs a route that is not recorded.

External participants especially. They did not agree to your tooling.

Where the recordings go

Cloud transcription sends the audio to a provider. For client conversations, legal or medical discussions, HR matters or anything confidential, that is a disclosure.

Check the terms: retention period, whether transcripts are used for training, who at the vendor can access them, where processing happens.

On-device transcription avoids the question and is now good enough for many purposes.

The retention problem nobody addresses

Meeting tools accumulate a complete, searchable, attributed archive of everything anyone said in every meeting for years.

That is discoverable in litigation. It is a target for an attacker. It is a record of every half-formed idea and frustrated remark anyone made.

Set a retention policy before the archive exists, not after. Most organisations have not, and are accumulating something they have not thought about.

A reasonable default: keep summaries and decisions long-term, delete transcripts and audio after a defined short period unless a specific reason exists to keep them.

Making the meeting summarise well

Most of the quality is determined in the meeting, not by the tool.

State decisions out loud, explicitly. "So we have decided X" produces a clean line the summariser cannot misread.

State owners out loud. "Sam is taking that." This single habit fixes the attribution problem, which is the tool's main failure.

Say when something is not decided. "We are not deciding this today" prevents a manufactured conclusion.

Recap in the last two minutes. Decisions, owners, deadlines. The summary will be built largely from this and it will be right.

Name people when you disagree with them, so the disagreement survives compression rather than being smoothed into consensus.

Five habits, no tooling, and they improve the record more than switching products would.