AI Summaries Capture What Was Said, Not What Was Decided

Automatic meeting summaries are genuinely useful and they have a specific blind spot, which happens to be exactly the thing anyone reads a meeting record for.

"The summary has everything except the decision."

What summaries get right

They are good at several things and it is worth being fair about them. They capture which topics came up and roughly how much time each took. They identify action items when somebody says "I will do X". They reflect tone reasonably well, including when a discussion was tense. And they are excellent memory aids for somebody who attended and wants to recall the shape of the conversation.

For that last use they are close to ideal, and the mistake is not using them. It is treating them as a record for people who were not there.

The one thing they miss

Read a summary of a meeting where a real choice was made. You will find something like: "The team discussed whether to include the CSV export in the March release. Concerns were raised about testing time. Alternatives including a partial export were considered."

Accurate, and it does not say what is being built. The reader, who was not there, still does not know whether to plan for the export. So they message somebody who was, and the summary has added a step rather than removed one.

Why the decision is invisible in the transcript

Decisions in real meetings are rarely declared. The sequence is usually: somebody proposes, somebody objects, the objection is partly answered, there is a pause, somebody says "yeah, let us do that for now", and the conversation moves on.

Nothing in that transcript is marked as the decision. "Let us do that for now" is textually weaker than the confident objection that preceded it, and the objection was longer and more articulate. A summariser weighing the text has no reliable signal that the short concluding phrase carried the authority.

Worse, it has no way to know who was entitled to decide. If the most junior person in the room said "let us do that", the transcript looks identical to the case where it was said by the person who owns the area, which is the distinction the reader most needs.

The false sense of coverage

The real cost is not the missing decision, which people work around. It is the belief that the meeting is documented.

Once a team has automatic summaries, nobody writes the decision, because the meeting is recorded and summarised and that feels like documentation. So the practice of stating the outcome in the last two minutes disappears, and the team ends up with more text and less retrievable decision than it had before. The same dynamic in written form is covered in meeting notes nobody reads.

What to do instead

Keep the summaries as a memory aid for attendees, and add two minutes at the end of the meeting where the person who made the call states the decision, the authority, and what was rejected. Four lines, written by them, in the place the absent colleagues will read.

The distinction underneath this is worth stating plainly, because it applies to more than meetings. A summary holds what a model heard. A record holds what a person wrote down and committed to. The second one can be relied on by a colleague who was not present, and the first one cannot, however well written it is.

That is the design StandIn takes. The decisions your team answers from are the ones people actually wrote down in their brief, in their own words, with a source under every answer, and it never guesses: when the decision is not in the record it says so and names who to ask, rather than reconstructing something plausible from the discussion around it. See how StandIn works.

Common Questions

Are AI meeting summaries reliable?

Reliable for topics, tone, and explicit action items. Unreliable for what was decided, because decisions in real meetings are usually implicit and carry no textual marker distinguishing them from the discussion around them.

Should we stop using meeting transcription?

No. Use it as a memory aid for people who attended, and add a human-written decision record for people who did not. The two artifacts do different jobs and only one of them is safe for a colleague to act on.

Who should write the decision if not the summariser?

The person who made the call, in the last two minutes of the meeting, read aloud. Reading it aloud doubles as a check that everyone in the room heard the same outcome, which is not always the case.

Why can a summary not identify the decision?

Because it would need to know who held the authority and which sentence carried it, and neither is present in the transcript. "Let us do that for now" from the area owner and from a new joiner look the same in text.