AVAY

Why AI Summaries Miss the One Line That Mattered

22 August 2026

Your AI summary missed the important part because summarization models rank sentences by how prominent they are — repeated, emphasized, spoken at length — not by how much they change what happens next. A decision said once, quietly, in six words loses every time to a paragraph of discussion that circles the same point. The fix isn't a better model; it's telling the tool what a decision sounds like.

A flow showing a transcript being chunked, scored for salience, and compressed into a summary that drops a quiet decision. segmented by… scored for sa… trimmed to le… compressed 1 Raw transcript everything said, unweight… 2 Chunking split into topic segments 3 Salience scoring ranks sentences by promin… 4 Compression keeps top-ranked, drops t… 5 Summary output reads clean, misses the q…
How a summarizer decides what survives compression

The one line that got left out

A 40-minute roadmap review produces a summary of six clean bullet points. Five of them cover the thing everyone talked about for twenty minutes — messaging, channel mix, the launch video that still needs a voiceover. The sixth line, the one that says the launch date actually moved to the 14th, isn't there. Someone said it once, seven words, in the middle of a sentence about something else, and it never surfaced again.

That's not a bug in the transcription. The words were captured; every syllable is in the transcript. The gap is in what the summarizer decided was worth keeping, and it made that decision the same way most abstractive summarizers do: by scoring sentences on how much of the meeting they represent, not on how much they change what happens Monday.

Why summarizers score by prominence

Most meeting summarizers, AVAY's included at the model level, rank sentences for salience before they compress anything. Salience is usually a mix of signals: how often a topic recurs across the call, how central a sentence is to its paragraph, whether other people echoed it, sometimes plain term frequency carried over from older extractive techniques like TextRank. None of those signals measure consequence.

A launch date changing is exactly the kind of statement that scores badly on every one of those axes. It's said once. Nobody echoes it, because everyone in the room already understood it and moved on. It's short, so it doesn't dominate a paragraph. By the metrics a summarizer actually uses, it looks like noise next to twenty minutes of marketing discussion — even though it's the only sentence in the meeting that required someone to update a plan.

What actually marks a decision

Discussion and decision don't sound as different as you'd expect, which is most of the problem. A decision is often flatter than the debate around it — plainer verbs, less hedging, said by whoever has the authority to say it and not contested afterward. 'We'll ship the 14th' carries no more emphasis than 'the 14th feels tight,' but only one of those sentences is binding.

What to ask a summarizer to do instead

The practical fix isn't a smarter model reading your mind. It's separating two jobs most summaries collapse into one: recapping what was discussed, and listing what was decided. Ask for those as two different outputs, scored on different criteria — one for coverage, one for commitment language and ownership — and the quiet sentence stops competing with the loud paragraph for the same three bullet points.

AVAY keeps a running decisions list during the call itself rather than generating one from the transcript afterward, on the theory that catching a commitment as it's said beats reconstructing it later. It still isn't foolproof: a decision phrased as a question, or one settled by a nod nobody says out loud, doesn't get logged either way.

This part isn't solved

No summarizer, AVAY's or anyone else's, reliably tells a hedge from a commitment when the wording is ambiguous. 'I guess we're doing that then' is a decision to everyone in the room and a maybe to a model reading it cold. Sarcasm, decisions made by silence, and commitments resolved outside the call — in a Slack thread an hour later — don't show up in any transcript-based tool, however it's built.

The honest advice is to treat the summary as a first pass, not a record. If a call produced a decision that matters, check it against the transcript or search for the sentence itself rather than trusting six bullet points to have caught it.

What gets ranked highWhat a decision usually looks like
RepetitionSaid multiple times, in different wordsOften said once, then never repeated
SpeakerComes from whoever talks the most in the meetingOften said by whoever has the authority to decide
PhrasingMarked with 'so', 'to summarize', 'in short'Marked with 'we'll', 'I'll', plain future tense
LengthLong explanations and repeated phrasing score higherOne clause, sometimes buried inside another sentence
What a length-optimized summary weighs versus what a decision actually looks like

Common questions

Can I fix this by asking for a longer summary?

Longer summaries include more raw content, but they don't reweight it — the missed line still loses to more-discussed material unless the summarizer is told to score for commitment, not volume. Length and coverage aren't the same axis as consequence.

Does asking for bullet points instead of prose help?

Not much. Bullet points reduce how compressed the text feels, but the sentences that become bullets are still the ones that scored highest on salience, so the same ranking problem carries straight through.

Why did the summary catch a throwaway joke but miss the actual decision?

Because the joke got a reaction — laughter, someone repeating it, a follow-up comment — which reads as salience to a model even though it changed nothing. The decision, said flatly and once, generated no such signal.

Is there any way to guarantee a decision gets captured?

No, not fully, on any tool built from a transcript. The best current practice is to say decisions explicitly — 'so, decision:' — or use a tool that logs decisions as they're said during the call rather than reconstructing them afterward, and then still check the transcript for anything that actually matters.

The short version

A summary built to be short will always lose the quiet, one-line decision to the loud paragraph around it — ask for decisions and discussion as separate lists, and check the transcript when it counts.

Read next

Try it on your next call

AVAY is a video meeting platform that transcribes the call itself — no bot joins, because there is nothing to join. Start one at avay.ai, read how each part works in the documentation, or see what it costs.