A week later the investigation starts
The call is over. Everybody nods. A week later it turns out the task was not done, and the investigation begins: who exactly was supposed to do it? The chat holds six versions of events, the calendar holds the meeting itself, and that is all. The recording, if there was one, sits there as an hour-long file nobody will open.
Usually one of two things happens next. Either somebody takes the blame to end the conversation, or the meeting is scheduled again — «let us sync». Both cost more than the twenty minutes for minutes that nobody could find.
The good news: turning a recording into minutes is exactly the kind of work a machine does well. The bad news: there are precisely two places where it errs predictably, and those two places are the ones you will have to check with your own eyes. Both are below.
Four different documents everybody calls «the minutes»
Half of the disappointment with «AI for meetings» grows from here: a person expected one thing and received another, and both are formally called minutes.
| What it is | What question it answers | Size for an hour-long meeting |
|---|---|---|
| Transcript | Who said what, verbatim | 8–10 thousand words. Raw material, not a document |
| Summary | What the conversation was about | A page. Easy to read, impossible to act on |
| Minutes | What was decided, who does it, by when | Half a page. Someone who missed the meeting can act on it |
| The outcome | What changed in the work after the conversation | Zero words: a task on the board, an event in the calendar, an entry in the knowledge base |
Note the direction: the lower the row, the fewer the words and the greater the value. The worth of minutes is not in volume at all — it is in how many actions they start per unit of text. Ten thousand words of transcript start none; five lines with owners and dates start five.
Hence the practical conclusion: if a tool produces «a brief summary of the meeting» and stops there, it has delivered row two. The hardest part is still ahead of it — and that is exactly the part usually left undone.
What working minutes are made of
Six blocks. Not five and not ten — each one closes a specific way of losing an agreement.
| Block | What is lost without it |
|---|---|
| Decisions | In a month you discuss what you already chose. And choose differently |
| Tasks | «I think we agreed to do something» — work that never started |
| Owners | A task with no name is never done: everybody sees it and nobody picks it up |
| Dates | «As soon as possible». The task is remembered on the day the result was needed |
| Commitments | Promises made outwards — to a client, a neighbouring team, a manager. Nobody asks you about them; they are recalled against you |
| Open questions | What stayed unresolved quietly turns into «we thought you had decided» |
Everything else — the date, the attendees, links to material — is framing. Useful, but not what decides whether the minutes work. How to word the items themselves so they read unambiguously we covered in the piece on meeting minutes and their template; this article is about something else — how to get those six blocks out of a recording rather than out of your memory.
Before and after
A chunk of an ordinary conversation — four minutes, four people. Familiar?
Anna: Look, we still have not settled the integration. I am for taking
the off-the-shelf connector, we have nobody to write our own.
Igor: Off-the-shelf means a licence at a thousand a month. I am not sure
finance will approve it.
Anna: Our own means two weeks minimum, and then maintenance.
Mark: Listen, how about I check the budget with Lena? If a thousand
goes through, the question is closed.
Igor: Well, if it goes through — no argument, the connector is faster.
Anna: Agreed then, we take the connector if the budget is fine.
Mark, can you find out by the end of the week?
Mark: Yep, I will write back by Friday.
Igor: And we need to tell the client the deadlines are not moving.
Anna: I will, on Thursday's call.
A hundred and fifty words, of which six matter. Here is the same thing once it has been worked through:
DECISION
Take the off-the-shelf connector instead of our own integration.
Condition: the 1000/month budget is approved.
Rejected alternative: build our own (≈2 weeks + maintenance).
Source: 00:14:20
TASKS
Check the 1000/month budget with Lena — Mark — by Friday (05 Sep)
Source: 00:16:05
OUTWARD COMMITMENT
Tell the client the deadlines are not moving — Anna — Thursday, on the call
Source: 00:17:40
OPEN QUESTION
What do we do if the budget is NOT approved? Discussed, not decided.
Three things in this example are worth noticing separately.
- The decision is recorded together with the rejected alternative. Two months from now somebody will ask, «why did we not build our own?» Minutes that contain only the chosen option cannot answer, and the discussion starts over.
- The decision is recorded together with its condition. «We take the connector» is false. «We take the connector if the budget goes through» is true. Conditional decisions break more plans than forgotten tasks, because everyone considers the matter closed.
- The open question is written out explicitly. Nobody voiced it as a question — it stayed as a hole in the conversation. And it is the one that fires if Lena says no.
Why minutes cannot be assembled in one pass
Here is a technical detail that explains why «summarise this in one request» produces something you cannot trust.
Take the conversation above and stop at Igor's «the connector is faster». Does that look like a decision? Very much so. Now imagine that twenty lines later somebody says, «oh wait, the connector does not support our protocol». A decision established from a fragment is wrong precisely because the fragment does not know the future.
Hence the split into two passes, with different things extracted by each.
| Pass | What it sees | What it can establish |
|---|---|---|
| Over fragments | A window of a few dozen lines, overlapping so a thought is not cut in half | Claims, proposals, questions, risks, problems, ideas, assignments and participants' positions — including «I agree, but only if», which is not the same as «I agree» |
| Over the whole meeting | All the findings of the first pass at once | Topics, decisions (with the rejected options), promises still standing at the end, and any agreement to meet again |
The rules without which this does not work:
- Every item must point at the line it came from. If it cannot, it is not produced at all. That is the only defence against beautiful minutes of a meeting that never happened.
- A proposal nobody adopted is not a decision. Neither is support voiced by one person. Neither is a promise later withdrawn.
- Every item carries a confidence. Stated literally — high; a vague hint — low, and such items are dropped rather than added to the minutes «just in case».
What the machine does not know about owners and dates
Now the honest part, and it matters more than anything else here. Owners and dates are the two places where automatic minutes err predictably. Not randomly, but by the construction of human conversation.
| What was said | How it looks to the machine | How it actually is |
|---|---|---|
| «I will take a look» | A promise by a specific person | A polite formula. There was no promise |
| «We should update the documentation» | Work named, no owner | Everybody assumed somebody else would do it |
| «By Friday» | The date of the nearest Friday | Sometimes the next one. Sometimes «by end of week, but ideally Tuesday» |
| «Mark, will you take it?» — «Mm-hm» | The agreement may be missing entirely: «mm-hm» often gets lost in noise | An owner has been assigned |
| «We will do it» | A promise by the team | A promise by whoever said it — or by nobody |
So a properly built tool distinguishes work from a promise: «we should update the documentation» is an assignment with no owner, while «I will update it by Friday» is a promise by a specific person. The first must not be attributed to anyone, and putting the name of whoever spoke loudest there is the worst thing you can do to minutes.
A practical rule worth adopting as a habit: decisions and task wording are read with your eyes, and owners and dates are always verified. That is thirty seconds per meeting — and precisely the thirty seconds the other twenty minutes of savings exist for.
Minutes are not the finish line. The outcome is
Go back to the first table, to its last row. Minutes that were read and closed are a document of intent. The work begins when their lines become objects in the systems where you actually work.
DECISION ───────▶ an entry in the knowledge base
(revisited months later: what we chose and why)
TASK ───────────▶ a card on the board
(title, due date, link to the minute of the conversation)
COMMITMENT ─────▶ a task with an addressee
(who it was promised to is half the meaning)
MEET AGAIN ─────▶ a calendar event
(but by hand only: invitations go to other people)
OPEN QUESTION ──▶ an item on the next meeting's agenda
And how this is done matters. The right form is not «the system created twenty tasks for you» but a proposal with the change shown: here is what will appear, here is what was there before, here is the quote from the conversation, apply or dismiss. The difference is fundamental: «move the due date to Friday» and «set the due date to Friday» look identical until the previous state is visible.
Separately — about what must never apply itself. A calendar event sends invitations to people outside your system; deletions and closures destroy things the system did not create. Consent to automate your own tasks is not consent to that.
How this works in Whisperer
Whisperer takes a conversation through to an outcome along the scheme described above — and where the text above says «a properly built tool», it describes exactly what is inside.
- A transcript with timecodes, split by speaker — from a live meeting on top of any service (no bot to invite, why that matters) or from an uploaded recording.
- Two passes of understanding, exactly those: over windows of the conversation — claims, questions, risks, assignments and participants' positions; over the whole meeting — topics, decisions with rejected options, promises still standing, and agreements to meet again.
- Every item with a quote and a link to the minute. An item that could not be attached to a line is not created.
- Notes by your template. Three built in (short summary, stand-up, task-focused) plus your own — up to twenty sections; minutes shaped to your team's habits are set up once and work from then on.
- Proposed changes rather than silent actions. A task on the board, an entry in the knowledge base, a calendar event — each is shown with the previous and the new state and a quote. By default everything waits for your confirmation; only creative operations on your own tasks may apply automatically, and only if you enabled that. Calendar, deletions and closures are manual under any settings.
- Reprocessing does not create duplicates. An existing task with the same title is not created again — instead a change of date is proposed if the dates diverge.
- Search across past meetings and export to txt, md, srt, vtt and json — the minutes travel to wherever your documents live.
The limits, plainly. The system does not assign owners for you: it distinguishes work with no owner from a promise by a specific participant, and a human puts the name on the card. It proposes dates from what was said out loud — and it will read «by Friday» as the nearest Friday, which is not always right. The free plan gives 60 minutes of recognition per month; meeting analysis, tasks and search are on paid plans. And no tool saves a meeting where nothing was decided: minutes cannot be extracted from a conversation without decisions, because they are not in there.
The wider picture of the category — what else this class of tools does — is in the pillar piece: AI meeting assistant.
Frequently asked questions
How do I make meeting minutes from a recording?
Three steps: get a transcript with timecodes, break it into decisions, tasks, promises and open questions, then lay the result out into a minutes template. The machine does the first two, the template does the third. What is left for a human is proofreading the wording and checking owners and dates.
How do minutes differ from a transcript and a summary?
A transcript answers «who said what», a summary answers «what the conversation was about», minutes answer «what was decided, who does it and by when». You cannot act on the first two: they contain neither owners nor dates. The value of minutes lies not in volume but in the number of actions they start.
Can AI determine the owner and the deadline itself?
Propose them — yes; guarantee them — no, and claiming otherwise would be dishonest. «I will take a look» reads to a machine like a promise, «we should update the documentation» like a task with no owner, and «by Friday» does not always mean the nearest Friday. A good system honestly distinguishes ownerless work from a specific person's promise, but the names and dates in important items are checked by a human.
What do I do with decisions that were conditional?
Record them with the condition: not «we take the connector» but «we take the connector if the budget is approved», plus a separate line for what happens if it is not. Conditional decisions break more plans than forgotten tasks, because everybody considers the matter closed.
Are tasks created in the tracker automatically?
By default no: a change is shown as a proposal — what will appear, what was there before, and which line it follows from. Automatic application is possible only for creative operations on your own tasks and only if you enabled it. Calendar events, deletions and closures stay manual always: their price is paid by more than the account owner.
Is this good enough for formal minutes?
As a draft — yes, and it saves most of the time. But a human signs the document, and for legally meaningful minutes the text must be proofread: a machine does not carry responsibility for a wording that goes out under your name.
What next
Minutes are not bureaucracy — they are the way not to hold the same meeting twice. Everything they need was already said out loud: the decisions, the promises, the dates. The tool's job is to keep them from dissolving between «everybody nodded» and «a week later it turned out».
Test that on your next call rather than on an archive: create an account — 60 minutes a month are free — turn recording on at your next meeting and look at what you have ten minutes after it ends. Compare it with what you remember: if the analysis found even one promise you would have forgotten, the tool has already paid for itself. And if the minutes come out empty — that is a useful result too: it means nothing was decided, and that is worth doing something about.