AI Meeting Notes for All-Hands and Company Meetings

The one meeting the whole company attends and nobody writes down

Think about who takes notes at your all-hands. In a one-on-one, both people are half-writing. In a customer call, someone owns the recap because a deal depends on it. In an engineering review, the person on the hook for the change is scribbling. But at an all-hands, the audience is three hundred people and the note-taker is nobody, because everyone assumes the size of the room means someone else has it covered. That assumption is wrong at exactly the moment it matters most.

The all-hands is where leadership says the things that set direction for the entire company at once: the number we are steering to this quarter, the priority that just moved up, the reorg that is coming, the hire that got approved, the answer to the question everyone was too nervous to ask until someone did. These are the highest-altitude commitments an organization makes. And they are the least reliably captured, because the meeting feels like a broadcast rather than a working session, and a broadcast does not feel like something you take minutes of.

This piece argues the counterintuitive half. The point of capturing an all-hands is not to help the people who were in the room. They heard it. The point is everyone and everything downstream of the room: the teammate in a time zone eight hours off who watched at double speed on Thursday, the person who joined next month and needs to know what was already promised, and the version of the company that has to settle, six months from now, an argument about what was actually decided. Without a record, all three run on memory, and memory at the scale of a whole company is not memory at all.

A recording is not a record

Most companies think they already solve this, because they hit record. The video exists, it lands in a shared drive, and the problem feels handled. It is not. A forty-five minute recording is not a record of what was decided; it is a haystack that happens to contain one. Nobody scrubs a timeline to relocate the sentence where the CEO committed to a launch date. The recording is technically complete and practically useless, because completeness without retrieval is just storage.

The gap is between having the meeting somewhere and being able to answer a question from it. Six weeks later someone asks, in a channel, whether leadership said the hiring freeze applied to backfills. The honest answer is that it is in the all-hands recording from three weeks ago, around minute nineteen, probably. Nobody goes and checks. Instead two people who were both in the room disagree confidently, and the org resolves it by whoever is more senior or more certain, which is how corporate amnesia works in practice: not a blank where the memory should be, but a confident wrong memory that nobody can cheaply overturn.

Atlassian's own guidance on asynchronous work makes the useful point that async communication inherently creates a record that can be referenced later, so people who were absent or heads-down can catch up on their own schedule (Atlassian, 'Asynchronous Communication Is the Backbone of Distributed Teams'). The all-hands recording is a record in that thin sense. The thing it is missing is retrieval: the ability to ask a plain question and get the exact spoken commitment back with the context around it, rather than a two-hundred-megabyte file and a suggestion to skim.

What actually needs capturing at an all-hands

The transcript of an all-hands is mostly narration. Slides get walked through, context gets set, someone tells a story about a customer. Very little of that is load-bearing. The load-bearing content is a small set of statements that will be quoted, disputed, or held against a person later, and those are what a capture system has to isolate.

First, commitments made out loud: dates, targets, and promises leadership stated to the whole company. 'We will ship the new billing flow before end of quarter' is a different kind of sentence than the paragraph around it, because it can be true or false in three months and people will remember whether it was said. Second, decisions and changes of direction: the priority that moved, the thing we are stopping, the bet we are now making. Third, numbers stated aloud, especially ones that differ from what people last heard, because a number said in an all-hands becomes the number the company believes until it is corrected in another all-hands. Fourth, and most valuable, the answers given in Q&A, which we will come back to.

A useful discipline, and one Atlassian recommends for meetings in general, is to capture not just what was said but the owner and the follow-up: end with the open questions and assign someone to each (Atlassian, 'How to run effective meetings in the era of hybrid work'). At an all-hands the owner is often leadership itself, which is precisely why it evaporates. A team owns its action items because a manager will ask. A company-wide commitment made from the stage has no manager asking, so unless it is written down as a commitment with a name and a date, it quietly becomes something everyone heard and nobody tracks.

The time zones make it worse, and the async catch-up makes it matter more

For a co-located team the all-hands is at least a shared moment; everyone heard the same words in the same room at the same time. For a distributed company that shared moment does not exist. Some people are live, some watch a recording the next morning, some read a summary in a channel, and some never get to any of it and absorb the decisions secondhand through their manager. The company does not hear the all-hands once; it hears fragments of it across a week, through different mediums, at different fidelities.

That is the environment in which a single authoritative record stops being a nice-to-have. When four people encountered the same all-hands four different ways, the only thing that lets them agree on what was said is a canonical text they can all point at. The distributed-work literature is blunt about this dependency: asynchronous communication works because it produces a durable reference, and it fails when the reference does not exist and people are left reconstructing a live moment they experienced only partially. A record that is searchable turns 'what did leadership say about X' from a game of telephone into a lookup.

There is a second-order effect worth naming. When the record is good, the pressure to attend live drops, which is the point of async done well: people stop burning a synchronous hour across incompatible time zones to receive information they could have retrieved. When the record is bad, the opposite happens. People feel they cannot miss the all-hands because missing it means losing the information permanently, so a distributed company drags everyone into a single clock to protect against its own forgetting. Good capture is what lets an all-hands be genuinely optional to attend live without being optional to know about.

The Q&A is the highest-value, least-captured part of the meeting

The prepared portion of an all-hands is the safest to lose, because it exists as slides and someone can reread them. The Q&A is the opposite: it is unscripted, it is where leadership says things they did not plan to say, and it exists nowhere but in the room and the recording. It is also where the most consequential sentences of the whole meeting tend to land, because a good question forces a real answer.

Someone asks whether the new policy applies to their team. Leadership answers on the spot, and that answer is now the policy for that team, even though it appears in no document, no slide, and no email. It lives in the ninety seconds of a recording that nobody will find, and it will be relitigated the first time it inconveniences someone. Capturing the Q&A as a structured list of question and answer pairs, tied to who asked and who answered, converts the most volatile part of the meeting into the most citable, which inverts the usual outcome where the prepared material is preserved and the actual commitments are lost.

This is also where botless capture earns its place. A meeting bot that joins an all-hands is a fourth kind of awkward: it broadcasts to the entire company that the session is being recorded by a third-party participant, it shows up in the attendee list, and it changes how candidly people ask questions. Capture that works from the screen and the audio on the presenter's or an attendee's own machine, with no separate participant joining the call, gets the same transcript and the same searchable record without turning the room's most valuable conversation into a performance for a visible recorder.

Failure modes you can predict

The promise nobody wrote down. Leadership commits to a date or a headcount from the stage. The company remembers a commitment was made but not the exact terms, so when it slips there is a diffuse sense of broken trust and no shared text to check the actual words against. A written commitment with a date is falsifiable and therefore fixable; a remembered one is just resentment.

The number that changed silently. A metric gets stated in Q2's all-hands, restated differently in Q3's, and nobody notices the two do not match because neither was captured as a durable figure anyone can compare. The company operates on a number it believes is current and is actually a quarter stale, which is the organizational-memory failure the concept of corporate amnesia names: not an empty slot but a confident, outdated belief that constrains decisions (Wikipedia, 'Corporate amnesia'). Capturing stated numbers over time makes the drift visible instead of invisible.

The 'we decided that at the last all-hands' dispute. Two teams build on incompatible readings of the same announcement. Each is sincere. The meeting was the source of both readings and the arbiter of neither, because the record is a video nobody will scrub. This is the failure the whole piece is about: the meeting that reaches the most people is the one whose decisions are hardest to pin down afterward, precisely because its size made everyone assume it was already documented.

How to run a captured all-hands

Start from an agenda that names the decisions and updates on it, not just the topics, so the capture has a spine to hang statements on. Atlassian's meeting guidance is to put the agenda in the invite and make items results-oriented; at all-hands scale that discipline is what separates a broadcast from a working session, and it gives you the categories a capture tool can sort statements into afterward.

During the meeting, let capture run passively rather than assigning a human scribe. A person taking notes at an all-hands either misses the discussion to write, or writes so little that the Q&A, the most valuable part, is the first thing dropped. Passive capture of the screen and audio produces the full transcript without pulling anyone out of the meeting, and the human effort moves to where it is actually useful: reviewing and confirming the extracted commitments afterward, which takes minutes.

Immediately after, produce the record the company actually needs: a short list of commitments with owners and dates, the decisions and direction changes, the numbers stated, and the Q&A as question-and-answer pairs. Post that as the canonical text, not the recording. The recording can exist as a backstop, but the thing you link people to, and the thing new joiners read, is the structured record. Then make it searchable, so the question that arrives six weeks later is a lookup instead of an argument.

What to look for in tooling at company scale

Retrieval over storage. The test is not whether the tool keeps a recording; anything does. The test is whether you can ask a plain question months later, 'what did we commit to on hiring at the June all-hands', and get the exact spoken commitment with its context, not a file to skim. If the answer requires a human to relocate the moment by hand, the tool has not solved the problem the recording already failed to solve.

Botless, no-participant capture. At all-hands scale a visible recording bot in the attendee list is a real cost: it signals surveillance to the whole company and dampens the Q&A. Prefer capture that runs from a screen and audio without joining the call as a participant, so the largest and most candid meeting you run is not the one you compromise to record.

Cross-meeting memory, not per-meeting notes. The value of all-hands capture compounds across quarters: the number stated in Q2 is only useful because you can set it against the one stated in Q3, and a commitment is only trackable if last quarter's is still findable this quarter. A tool that produces a good summary of one meeting and forgets it is a notepad. What an all-hands needs is durable, searchable work memory that lets statements across meetings be compared, which is a different product than a per-call summarizer.

Privacy and control, because the whole company is in the room

An all-hands is the most broadly attended meeting a company runs, which raises the stakes on who can reach the record and how long it lives. The right default is that capture is deliberate rather than ambient: nothing joins the call, nothing auto-records in the background, and the person capturing is doing so on purpose, not because a tool is always listening. That keeps the record something the company chose to make rather than a byproduct of software running unattended.

Control also means the record is disposable when it should be. Audio does not need to be retained to produce the text; transcription can happen in real time and the audio be discarded, so what remains is the searchable record and not an archive of everyone's voice. Individual notes should be removable immediately, and a full account deletion should actually remove the underlying data on a bounded timeline rather than leaving copies behind. None of this is a certification or a compliance claim; it is the baseline posture a system should have before it is trusted with the one meeting the entire company can see.