How to catch what AI meeting minutes make up

Saegyeo guide · October 9, 2026

Summaries produce plausible sentences

Ask an AI to summarize a meeting and you get readable sentences. The trouble is that the summary alone cannot tell you whether a sentence was actually said. “Let’s revisit the budget later” can turn into “Budget to be split 70:30” and still read naturally.

Minutes are what people trust at the next meeting, so one such line becomes a promise or a date. Every line of AI-written minutes needs a way back to its evidence.

Step 1. Require a quote on every line

For each decision, action item and issue, have the AI copy the utterance it rests on, word for word. With the speaker and time attached, you can jump straight to that part of the recording.

Asking for quotes is not enough on its own, because an AI can make up a plausible quote too.

Step 2. Match quotes against the transcript, character by character

Code, not the AI, checks that each quote really appears in the transcript. Ignoring spaces and punctuation, if a single character differs, the line is dropped. It is a plain comparison with no judgement in it, so the same input always gives the same result.

Step 3. Summarize several times and keep what repeats

Summarize the same meeting three times independently and keep only items that appear at least twice. An item that shows up once may be chance. A fair share of candidates drop out here, and some of them may have been right, which is worth knowing.

Step 4. Keep uncertain items apart

When the evidence is unclear, mark it “none”; when the three passes disagree, mark it “uncertain”. This keeps doubtful items from blending into the summary and reading as settled.

What this does not catch

A correct quote does not make the reading correct. If the minutes quote “we’ll look into it” and write “decided to review”, the quote is in the transcript but the judgement is wrong. That is why a tap on a line’s time should play the utterance from there.

If the transcript is wrong, the quote check is measured against the wrong transcript. For lines with names, numbers or amounts, a quick listen to the recording is the safe choice.

What we measured in Saegyeo

Seven of the founder’s real meeting recordings, each summarized three times, 21 passes in all (counted Oct 8, 2026).

0 · lines dropped because the quote was not in the transcript
30% · candidates dropped for appearing in only 1 of 3 passes (100 of 333)
233 · items kept in the minutes (decisions, action items, issues, objections…)

Seven meetings is a small sample, there is no human-labelled answer key yet, and the transcripts were not checked by a person. Next to measure: decision precision (extracted decisions that are real), decision recall (real decisions not missed), and how often the quote truly supports the item.

What to check in any tool

Does each line of the minutes carry a supporting quote or a recording time?
Does a person or code check that the quote is actually in the transcript?
Are phrases like “we’ll look into it” written up as decisions?
Do action items have an owner and a deadline that someone actually said?