The meeting summary is not the meeting

AI meeting summaries are useful because they leave things out. The managerial risk arises when the shortened account becomes the only practical record available to later readers. Checking that its statements are supported is not the same as checking that decision-relevant material survived. For consequential meetings, I would preserve a proportionate route back to the underlying record, review omissions as well as assertions, and keep subsequent analysis separate from what participants actually said or agreed.

When I first began discussing the publishing system behind Andrew’s Intel, a question about meeting summaries bothered me. How would I know what an AI-generated account had left out? I had been in the conversation. Someone who had not attended would have even less chance of spotting a missing point.

A Facebook reflection by Genny Harrison brought that concern into sharper focus. Writing about the 2011 film Margin Call and the financial crisis, she examines how complex information is compressed as it travels towards people with authority. Somebody—or something—selects what survives. Her observation helped me connect the organisational problem with the apparently smaller business of summarising a meeting.

My concern is not that summaries should contain everything. A summary that did would have failed at its job. It is that the convenience of the short version can conceal a second decision: which parts of the conversation the next reader will have a reasonable chance of knowing about.

This is already a practical question, not a prediction about future AI. Microsoft’s Teams documentation describes AI-generated meeting notes and follow-up tasks, alongside ways of navigating the underlying recording. It also explicitly warns that generated content can be incomplete. Those facilities can help; whether a particular reader can meaningfully check the account remains the important question.

Accuracy is only half the test

Consider an illustrative supplier-review meeting. Most of the discussion concerns price, the timetable and a possible pilot. Someone briefly asks whether the organisation could retrieve its data in a usable form if it later changed supplier. Nobody has tested that. The conversation moves on.

A summary reports that the cheaper provider was favoured, a pilot was discussed and procurement will seek revised terms. Every statement might be supported by the meeting. The unanswered question about leaving the supplier is nevertheless absent.

An executive reading the summary can check those three statements without ever discovering the fourth issue. The omission does not have to make any individual sentence false to make the account less useful for the decision now being taken.

I would distinguish factual fidelity—whether the account is supported by the source—from decision-relevant coverage—whether the account preserves what this reader needs to judge the matter. Neither requires a transcript-length summary. They require different checks.

There are also less clean cases. A condition can disappear from an apparent commitment, or a proposal can be presented as agreement. Those are distortions as well as omissions. The more subtle problem is the accurate but insufficient account: nothing obviously wrong, but something important missing.

Research recognises omission as a distinct problem. In a 2025 study, Frederic Kirstein and colleagues assembled 200 automatically generated meeting summaries with human annotations covering nine error types, including omissions. A separate identification-and-refinement process improved summary quality in their experiments. That supports testing for missing material, not just invented material. It is not a failure-rate estimate for today’s commercial meeting assistants.

Important to whom—and when?

It is tempting to solve this by writing a better instruction: retain the important points, preserve disagreement, identify risks.

Those are sensible instructions. They do not settle what importance means.

The procurement reader may care about exit arrangements. The project manager may need dependencies. A colleague joining three months later may want to understand why an alternative was rejected. A promising idea that occupied thirty seconds might matter more to a future discussion than the timetable that occupied twenty minutes.

The QMSum research benchmark makes the reader’s question part of the task: select relevant passages from a meeting and summarise them in response to a query. Its design acknowledges that different users need different accounts of the same discussion. The 2021 paper is useful here for that distinction, not as a verdict on current model capability.

There is still a limit. We can specify today’s questions better than tomorrow’s. I cannot reliably instruct a summariser to preserve every observation whose importance I have not yet recognised. Asking it to retain “all potentially useful details” largely returns us to the original material.

That is why I would not make the prompt carry the whole burden. The surrounding process needs to allow later readers to ask a different question of the source, rather than only a different question of its summary.

Keeping a transcript is not the whole answer

A transcript provides a route back, but only if it exists, remains available to the appropriate people and can be searched or navigated without unreasonable effort.

Those conditions should not be assumed from the existence of a summary. Zoom’s published instructions, for example, show an optional “Retain meeting transcript” control when starting meeting summaries on mobile. Summary generation and the availability of a retained transcript are choices to check separately. This is a documented workflow detail, not a claim that every Zoom account has the same configuration.

Nor is the transcript the meeting itself. A captured record may contain transcription errors, miss overlapping speech or omit material discussed outside the recording. A pause does not prove reluctance; silence does not establish consent. No later analysis should invent the missing evidence.

I am not arguing for recording every conversation or retaining everything indefinitely. Sensitive discussions may need restricted records, limited retention or no recording at all. The appropriate choices belong to the organisation’s privacy and records-governance process. Where the underlying material will not be kept, that limitation should shape how much authority the summary is allowed to carry.

For a consequential decision, a short, explicitly confirmed decision record may be more appropriate than a general recap. It should distinguish what was agreed, the conditions attached and what remains open. A summary can help prepare that record; it should not quietly become one simply because it looks finished.

Check in both directions

My preferred test would run in two directions.

First, work from the summary back to the source. Is this statement supported? Was this actually agreed? Does the cited passage include a qualification that changes the meaning?

Then work from the source towards the summary. Which unresolved questions, objections, conditions and alternative proposals deserve representation for this use? Where do they appear in the short account—or why have they been excluded?

That second pass needs access to the underlying material. Asking an assistant to identify omissions while giving it only its own summary cannot establish what the meeting contained.

An agenda or topic index linked to passages can help a reviewer see the ground covered. Links beside included claims serve a different purpose: they help check those claims, but cannot by themselves reveal an entire topic that never entered the summary. Both forms of access are useful; neither is a completeness guarantee.

For low-stakes catch-ups, a quick participant correction may be proportionate. Before a summary influences a substantial commitment, I would want a named reviewer and a more deliberate check of decision-relevant omissions. The aim is not to make everyone reread every transcript. It is to make challenge possible where the consequences justify it.

Expansion needs its own boundary

This concern also influenced the meeting-analysis approach I have been developing. I want it to do more than make a conversation shorter. Its brief includes expanding useful points: examining an unfinished argument, connecting related observations and developing possible implications.

That can be valuable, but it introduces the opposite danger. Compression may remove something that was said; expansion may add something that was not.

I would therefore keep three kinds of material visibly separate. The record says what participants stated or agreed, with a route to the supporting passage. Analysis offers an interpretation of that material. A proposal develops a possible next step or a new combination of ideas for consideration.

In the supplier example, the record could say that data portability remained untested. Analysis might explain the possible implications for switching costs. A proposal might recommend testing an export before committing. The third step could be useful without having been discussed or approved in the meeting.

A system that develops a compelling new option has not discovered a hidden decision. It has produced something new for people to judge. I would rather preserve that distinction than make the output read as though the conversation had been more complete or more conclusive than it was.

Before the summary becomes the record

For a meeting whose output will influence a significant decision, these are the questions I would take to the team responsible for the workflow:

On a narrow screen, each question is shown as a separate card.

A useful supplier demonstration would show how a reviewer finds a deliberately omitted issue in a sample meeting, not merely how quickly the product generates an attractive recap. I would also ask what becomes of source links when access permissions change or the underlying record expires.

A shorter route, not the only route

Human-written minutes also select and omit. This is not a demand that AI meet an impossible standard of perfect preservation while people receive a free pass. It is a reason to be explicit about the authority of any compressed account.

I want AI to reduce the work of finding and understanding useful material. I do not want the only convenient account to become unchallengeable because the evidence behind it is inaccessible, or because nobody is responsible for checking what disappeared.

The meeting summary should be a shorter route into the conversation—not the only route left.

Useful distinctions

Factual fidelity: whether statements in the account are supported by the source. Decision-relevant coverage: whether the account preserves what matters for its intended use. Provenance: where a statement came from, including whether it was recorded, inferred or proposed afterwards.