What a Meeting Summary Does to Disagreement

The missing condition

Consider a hypothetical product meeting. A manager proposes shipping on Friday. An engineer says Friday is possible if the remaining tests pass. A colleague asks whether customer support will be ready. The manager asks for an update tomorrow, and the discussion moves on. A recap says the team agreed to a Friday launch, with engineering and support preparing delivery.

Every subject appears: Friday, tests, support, preparation. Yet the recap has supplied the agreement that the conversation never reached. A more faithful version would say that Friday was proposed, testing remained a condition, support readiness was unresolved, and an update was requested. This example is invented; it illustrates a distinction, not a documented product failure.

The distinction is between remembering what a meeting discussed and preserving what its participants committed to. Our argument is that a useful summary should be judged by both. A reader deciding whether to start work needs the status of a statement as much as its subject.

Coverage and decision status

The 2021 QMSum paper introduced query-based meeting summarization: selecting and summarizing discussion relevant to a question. Its annotations include speaker opinions and reasons behind proposals, alongside broader discussion summaries. The authors also distinguish word-overlap evaluation from factual consistency and relevance to the question. This benchmark offers a useful way to ask more of a recap than whether it mentions the right topics.

For our purposes, “What did the team discuss?” and “What did the team decide?” require different answers. A response can be accurate for the first question and misleading for the second. Even a true sentence about a proposal becomes misleading beneath a heading that labels it an adopted decision. The heading participates in the claim.

This separation has precedents in human annotation. The AMI corpus documentation describes summary sections for an abstract, decisions, problems or issues, and actions. Annotators linked summary sentences to supporting dialogue acts. The page does not provide reliability results for that summary scheme, so the structure should not be mistaken for evidence that classification is effortless or universally agreed.

Our proposal extends that separation to an everyday reading test: could someone distinguish an idea being explored from an instruction they are expected to carry out? A short recap that passes this test can be more useful than a comprehensive narrative that blurs those categories. Brevity should come from removing repetition, not from removing the condition that determines whether work should begin.

Who said it matters

In a 2022 dialogue-summarization study, Zhichao Geng and colleagues examined speaker confusion and missing speakers in BART-generated summaries. Their speaker-aware training improved factual consistency in evaluations on SAMSum and AMI. This is evidence that attribution is a distinct, tractable problem. It is also bounded evidence: the study used older models, selected dialogues for human evaluation, and truncated inputs. Its results do not estimate error rates for current commercial meeting assistants.

Attribution matters because an objection can change meaning when it changes owners. In a hypothetical design discussion, the person proposing a deadline might repeat a colleague's concern before answering it. Recording that repetition as the proposer's own objection would distort the exchange. Writing that “the team was concerned” might erase a meaningful difference between a single reservation and a shared constraint.

Names should carry a purpose. A summary need not attach a person to every passing preference. It should preserve attribution when it establishes who accepted a task, who raised an unresolved question, or whose position is being represented. When the speaker is uncertain, leaving ownership unconfirmed is more useful than confidently allocating work to the wrong person.

Disagreement changes during a meeting

The AMI decision-discussion instructions, dated June 18, 2007, explicitly allow one decision to span separated parts of a meeting. They also distinguish decisions originating outside the meeting from recaps of earlier decisions. These are annotation instructions for a particular corpus, not a general test of organizational consent. They nevertheless make a useful distinction visible: discussion location and decision origin are separate facts.

That matters when checking dissent. Finding an objection near the beginning does not establish that it remained unresolved at the end. A later answer may satisfy the objector; a later decision may proceed despite the objection. Conversely, a confident final recap from one participant does not by itself demonstrate that everyone changed their position.

Preserving disagreement should therefore include its disposition. Was the concern answered, withdrawn, left open, or acknowledged while a decision proceeded? These are our proposed reading categories. They protect against both false consensus and false deadlock. A dissenting participant does not automatically hold a veto, and a decision need not imply unanimity.

Where the conversation supplies no answer, the summary should preserve that limit. “No resolution recorded” describes the evidence available. “The team rejected the concern” claims an event. The first should not silently become the second simply because the meeting ended or the agenda advanced.

A recap someone can act on

We propose organizing an operational recap around the state of the work. Put adopted decisions together. Separate proposed actions from accepted assignments. Keep unresolved questions beside the decisions or proposals they affect, so readers encounter the uncertainty before acting. This is a design recommendation, not a format validated by the cited studies.

For the hypothetical Friday discussion, the useful entry would be compact: launch date proposed; testing condition outstanding; support readiness unconfirmed; update requested for tomorrow. It should not name an owner for that update unless the conversation establishes one. Empty fields can reveal an unfinished meeting rather than a deficient summary.

A link to a transcript passage helps only if the passage supports the status claimed. A sentence proposing Friday supports “Friday proposed.” It does not support “Friday agreed.” Where later discussion changes the outcome, the supporting material should include that change. A convenient citation to an earlier mention is insufficient.

This approach also separates genuine uncertainty from weak transcription. A participant may be uncertain about readiness, or the system may be uncertain about what the participant said. Those problems require different follow-up: resolving the readiness question or checking the recording. Combining them into a vague confidence warning tells the reader too little.

Review the verbs

A practical review can start with verbs such as agreed, approved, assigned, rejected, and confirmed. For each, inspect what in the conversation justifies it. Then check whether conditions and unresolved objections survived. This directs attention toward the parts of a recap that could change someone's next action.

A local evaluation could use clearly labeled hypothetical variants of the same discussion: one ends with agreement, another with deferral, another with a decision despite a reservation. A useful summarizer should produce different decision records even when the vocabulary barely changes. Include cases where an early objection is resolved, so the evaluation does not reward mechanically preserving every disagreement forever.

Such a test would supplement, rather than reproduce, the historical studies cited here. It would still require people to judge ambiguous exchanges. When they disagree, that disagreement is information about the source or the evaluation rule; forcing a single tidy answer may conceal the very problem being tested.

Meeting summaries earn trust when they make the next step clear without inventing permission to take it. Sometimes the most consequential thing to preserve is that a question remains open—and exactly what would have to happen for it to close.

Sources

• Zhong and colleagues, QMSum: A New Benchmark for Query-based Multi-domain Meeting Summarization. NAACL, June 2021.
• AMI project, AMI Corpus: Annotation. Undated documentation.
• Geng and colleagues, Improving Abstractive Dialogue Summarization with Speaker-Aware Supervised Contrastive Learning. COLING, October 2022.
• AMI project, Coding Instructions for Decision Discussion Segmentation of the AMI Meeting Corpus. Version 2.3, June 18, 2007.

Primary sources consulted October 2, 2026. Historical research informs the analysis; no current product performance is claimed.

Related reading

• The Meeting Bot Becomes Corporate Memory — capture, retention, and the institutional record.
• The Memory Summary Loses the Clock — what compression can remove from chronology.
• The Notification Summary Becomes the Attention Clerk — interpretation at the point of attention.

Production: commissioned by the site operator; researched and drafted by GPT-6 Astra; editorial and source review by the coordinating AI assistant.

Return to Blog