How to Extract Decisions From Meeting Transcripts (Not Just Summarize Them)

The migration meeting ended at 3:30 on a Tuesday in March. The recording bot was in the call, so there is a transcript — about 9,000 words of it, which is roughly what an hour of overlapping speech produces — and a tidy AI summary with highlights and action items. Everything about that meeting is, in some sense, on file.

Six months later a new engineering lead asks the obvious question: "Why didn't we just defer this to Q4?" And nobody can answer. The person who made the case against deferring has left. The argument was real — it was raised in three separate asides, forty minutes apart, and it carried the room — but it appears in no summary, because summaries keep highlights, and reasoning that arrives in fragments is never a highlight.

The transcript is the meeting. It is not the decision. This post is about the difference — what a decision record actually contains, why pulling one out of a transcript by hand eats an afternoon, and how AI extraction changes the economics when it targets the decision structure instead of the recap.

A transcript is the meeting. It is not the decision.

The transcript-and-summary problem is essentially solved. Half a dozen tools will sit in your call, produce a clean transcript, and hand you a summary with the highlights and the action items. That is genuinely useful, and you should keep using it. But notice what it gives you: a compressed retelling of what was said. That is a summary's job, and it does it well.

What it does not give you is a record of what you decided — the choice the group committed to, the alternatives you ruled out, and the reasons you ruled them out. No amount of better summarization produces that, because the two documents answer different questions. A summary answers "what happened in this meeting?" A decision record answers "what did we settle, and why did we reject the other options?"

The distinction matters most at the moment it is most expensive: months later, when the question reopens. A summary is silent on "why didn't we defer?" — the argument against deferring was said out loud, then compressed away. A decision record answers it in a sentence. (For the closely related trap of logging tasks instead of the choices behind them, see action items vs. decisions — and for what happens to all that uncaptured reasoning when people leave, see the institutional knowledge crisis.)

Researchers have been trying to pull decisions out of meetings for twenty years

This is not a problem AI notetakers discovered last year. In 2005, the AMI project recorded and annotated roughly 100 hours of meetings precisely so researchers could study what actually happens in them, and by 2007–2008 groups were building systems to detect decision points in that corpus — Hsueh and Moore's decision-detection work, and Fernández and colleagues' models of decision sub-dialogues among them.

The consistent, humbling finding from that line of research: decisions in real meetings are negotiated, scattered, and implicit. People rarely say "we have decided X for reasons Y and Z." They circle a topic, drop it, return to it, half-commit, and move on — which is exactly why decisions are what both hurried humans and summarization models lose. The reasoning you need is distributed across turns that no single highlight contains.

What to actually pull out of a transcript

Extracting decisions is not "summarize, but shorter." It is a targeted read for four specific things. Miss any one and you have a note, not a record.

Software engineering institutionalized this exact shape in 2011, when Michael Nygard proposed architecture decision records: short documents capturing the context, the decision, and its consequences, kept with the code they explain. ADRs spread because they solved the same reopening problem — a new engineer asking "why is it built this way?" gets an answer instead of an archaeology project. Put the four items above together and you have the general-purpose version: the spine of a decision audit trail that survives the people who were in the room.

The manual method — and why it drains an afternoon

You can extract all of this by hand, and the method is straightforward:

1. Read the transcript end to end, marking every turn where a choice, an objection, or an alternative appears.
2. For each decision, write the resolution as a single sentence — the thing the group committed to.
3. Scroll back through the discussion and collect the options that were floated and rejected, so the choice has visible alternatives.
4. Reconstruct the arguments — quote or paraphrase who argued for and against each option, and attach the reasoning to the right side.
5. Name the owner, and file the whole thing somewhere it can actually be found again.

It works. It is also the part everyone quietly skips, because a real transcript is thousands of words of overlapping speech, tangents, and half-finished thoughts — and as the meeting-speech research above predicts, the reasoning you need is scattered across it, often never stated cleanly in one place. Reassembling a con that was raised in three separate asides forty minutes apart is slow, careful work. Do it well and it eats an afternoon per meeting. Do it in a hurry and you get a summary again — which is exactly the thing that was not enough. The tedium is not a discipline problem; it is a structure problem, and structure is automatable.

"But my AI notetaker already lists the decisions"

Fair objection — most meeting tools now emit a "Decisions made" bullet list, and if all you need is topics plus action items, that summary is genuinely enough. But there are two problems with leaning on it as your decision record, and they are different kinds of problem. (For the wider landscape — which kind of AI tool fits which decision job — see our honest roundup of AI decision-making tools.)

The first is reliability. When researchers had humans annotate 200 machine-generated meeting summaries for the QMSum Mistake dataset, omission — significant decisions and actions missing outright, or mentioned without their substance — showed up as one of nine recurring error classes, alongside hallucinated specifics that were never said (Kirstein, Ruas & Gipp, COLING 2025). Summarization models are optimized for coverage of what was discussed, not for guaranteed capture of what was settled — and the scattered, implicit way real decisions are reached is precisely the hard case.

The second problem is there even when the bullet is correct. "Decided: ship the migration in Q3" is the resolution alone — the options considered and the arguments on both sides are still gone, and those are the parts you need when the decision is questioned later. A correct one-line bullet is a note. The record is the structure around it.

How Argumentree extracts the decision, not the summary

Argumentree reads the transcript for a different target. Instead of compressing what was said into a narrative, its AI extraction looks for the decision structure: the question under discussion, the options weighed against each other, and the arguments raised for and against each one. It assembles those into a pro/con argument map — the same shape a good facilitator would build on a whiteboard — with a draft decision and a suggested owner attached.

The output is not another summary living next to your transcript. It is a structured, queryable decision record with the reasoning still attached and a full audit trail — the choice, its alternatives, and the arguments on both sides one click away the next time the question comes back. If turning meetings into defensible decisions is a recurring job for your team, this is the heart of what Argumentree calls meeting intelligence.

The transcript is the source. The record is the asset.

Keep your notetaker — the transcript it produces is the raw material everything else depends on, and its summary is a fine convenience for people who missed the call. Just stop asking it to be something it is not. A compression of what was said cannot answer what was settled and why, and the moment you need that answer is precisely the moment the people who could reconstruct it from memory are gone.

Extraction — by hand on your highest-stakes meetings, or automated with a human in the loop everywhere else — is what turns a pile of recordings into an asset that compounds: decisions with visible alternatives, reasoning that survives departures, and reopened questions that get answered by a link instead of a rerun.