A promise to remember the decisions
Most meetings end not with decisions but with a promise to remember the decisions. «Thanks everyone, we are agreed» — and seven people leave with seven different versions of what was agreed. From then on the thing that works is not a system but memory: whoever has the better one is right a week later.
The funny part is that all the links were there. The task was voiced. The person was named. The date was said out loud. What gets lost between «said» and «done» is not information — it is the transition: nobody moved what was said into the place people look at every day.
Below: what that transition consists of, where it breaks for different teams, and what a machine can realistically do about it. With the caveat that removes half the expectations straight away: it is good at finding and sorting, and it does not work for you.
Six links, and any of them can break
The path from a conversation to a finished thing looks like this:
TALK ─▶ AGREEMENT ─▶ TASK ─▶ OWNER ─▶ DUE DATE ─▶ DONE
│ │ │ │ │ │
said recognised written a name a date remembered
out loud as a into a attached named in time
promise system
Each link breaks in its own way, and each is fixed differently.
| Link | How it breaks | What happens next |
|---|---|---|
| Talk → agreement | «I will take a look» sounds like a promise but was not one. «We should update this» names work with no owner | Half the task list is polite formulas, and half the real promises never made it in |
| Agreement → task | The promise lives in a head and a chat rather than in the system people look at | «I think we agreed something» a week later |
| Task → owner | The task exists, the name does not | Everybody sees it and nobody does it. The most common break of the six |
| Owner → due date | «As soon as possible» | The task surfaces on the day the result was already needed |
| Due date → done | The date is written where nobody looks | You remember on Monday about Friday |
Notice: not one of these links is about intelligence or discipline. They are all transfers of information from one place to another — exactly the work humans get wrong and machines do not.
Why a transcript does not solve it
The first answer that comes to mind is «let us record everything». So you record. You get an archive of hour-long transcripts, and nothing changes.
- Nobody reads the archive. To find a promise in a transcript you have to know it is there. But the whole problem is that you forgot about it.
- Search needs the word. «Sergey analytics» will be found if Sergey said the word «analytics». He said «I will look at the funnel numbers» — and nothing is found.
- A transcript answers a different question. It is about «what was said». You have to work from «what we do». Those are different documents, and the second does not appear from the first on its own.
- Too much was said. An hour-long meeting contains two or three dozen phrases that resemble an assignment. Three or four of them are real promises.
So «we turned recording on» is not the same as «we stopped losing tasks». Recording is a precondition, not a solution.
«AI recorded the meeting» and «AI understood what has to be done»
The difference between those two is the entire meaning of the phrase «AI meeting analysis», which in product descriptions can mean anything at all.
| AI recorded the meeting | AI understood what has to be done | |
|---|---|---|
| Output | Text. Sometimes with a summary | A list of objects: task, promise, decision, open question |
| Does it tell kinds apart | No. An idea, a proposal and a promise are identical sentences of text | Yes, and that is the crux: an idea is not a task, a proposal is not a decision, «I will look» is not a promise |
| Is there an address | No. To act you must read it and move it by hand | Yes: a card on the board, a date, a link to the minute of the conversation |
| Is it checkable | Only as a whole: re-read an hour | Item by item: each carries the quote it was derived from |
| What the human does | Everything | Checks names and dates, dismisses the extras |
The key row is the second. The ability not to make a task out of an idea matters more than the ability to make one out of a promise: a list where half the items are somebody else's thinking out loud gets closed and never opened again.
One sentence under a microscope
Take a typical line from a product meeting:
«Let us update the onboarding then, and Sergey will look at the analytics by Friday.»
One sentence. Sixteen words. Inside it are two tasks with different status, and confusing them is not allowed.
TASK 1 Update the onboarding
owner NOT NAMED — «let us» is addressed to nobody
due not named
status the work is acknowledged as needed; who and when is for people
source 00:23:10
TASK 2 Look at the analytics
owner Sergey — named directly
due by Friday
status a promise by a specific person
source 00:23:14
CONTEXT both tasks grew out of a discussion of the conversion drop
at the second step of signup (lines 00:21:40–00:23:20)
Three things matter most here.
- The first task stayed without an owner — and that is the correct answer. Putting there the name of whoever proposed it, or whoever spoke loudest, is the worst move available. A task with no name, honestly marked as ownerless, is fixed in ten seconds; a task with the wrong name is never fixed, because it looks plausible.
- «By Friday» becomes a date — but it is worth checking. The nearest Friday is right almost always, and wrong exactly when the meeting runs on a Thursday evening.
- Context matters more than wording. «Update the onboarding» is unclear in two weeks even to its author. «Update the onboarding — conversion drops at the second step» is clear two months later.
Six teams, six different lists
«Tasks after a meeting» means different things in different departments. It is worth looking at what exactly each of them loses.
A product team. What gets lost is not tasks but conditions: «we do A if the data confirms the hypothesis». A month later everybody remembers «we do A» and nobody remembers «if». What has to be caught here are conditional decisions and open questions: they are the real output of a product meeting.
Engineering. What gets lost is the technical reasoning: everybody remembers «we move it to a queue», nobody remembers «because at peak load the synchronous call saturates the pool» — and six months later the architectural decision is discussed again. A detailed treatment is in AI for developer meetings.
Sales. What gets lost are promises to the client — the most expensive kind. «We will send the quote by Wednesday», «we will check about their CRM integration». Nobody asks you about them; they are recalled against you, and the price is the deal. Here it matters who it was promised to, not only what was promised.
An agency. What gets lost is the scope boundary: somebody said «and could you also fix this bit» — and that is either a task inside the contract or grounds for a change order. An unrecorded «could you also» becomes «you promised» three weeks later.
Management. What gets lost are the manager's own promises: «I will find out about the budget», «I will talk to legal». Nobody checks those — which is exactly why they do not happen, and the team concludes that agreements only run one way.
One-on-one. Everything gets lost: the meeting is not minuted because it is personal. And that is where the most important promises live — about growth, salary, a change of role. Three months later the two sides remember different things, and the conversation ends in «we discussed this» versus «we did not decide this».
What happens to a found task next
This is where tools start promising too much. Let us be plain: fully autonomous task management does not exist, neither here nor at the neighbours — and it is a good thing it does not. A system that creates, moves and closes tasks by itself based on what it heard in a conversation errs silently, and people clean up afterwards.
The right form is a proposal with the change shown:
- You see what will appear and what was there. «Move the date to Friday» and «set the date to Friday» look identical until the previous state is shown.
- You see where it came from — a quote and the minute of the conversation. An item that could not be attached to a line is not created at all.
- A human applies it. By default everything waits for confirmation. The only thing reasonable to apply automatically is creating your own tasks — and only if you enabled it.
- Some things must never be automated. A calendar event sends invitations to outsiders; deletions and closures destroy what the system did not create. Consent to automate your own tasks is not consent to that.
- Reprocessing does not create duplicates. A task with the same title is not created twice — a change of date is proposed if the dates diverge.
And the last link — «remember in time». A task lying on the board can still go undone; the difference appears when the approaching deadline is raised with context: not «you have a task» but «you promised to look at the analytics by Friday, the promise comes from Tuesday's meeting, status open».
How this works in Whisperer
Whisperer breaks a conversation down into objects rather than into a retelling. What happens after a meeting:
- Analysis in two passes. Over fragments it extracts assignments, proposals, questions, risks and participants' positions; over the whole meeting — decisions, promises still standing at the end, and agreements to meet again. A fragment may not declare a decision: what sounds settled is overturned twenty lines later.
- Work and promise are told apart. «We should update the onboarding» is an assignment with no owner; «I will look at the analytics by Friday» is a promise by a specific person with a date. A promise also keeps who it was made to.
- Weak findings are dropped. Every item carries a confidence; vague hints never reach the list, and the list itself is capped — twenty cards a person will work through, fifty they will close unread.
- Every item comes with a quote and a minute. Checking takes seconds instead of re-reading an hour.
- Tasks reach the tracker board as proposals — with a title, a due date and a link to the source. Decisions do not become tasks: «we chose option A» is knowledge of the meeting, not an assignment, and turning it into work means inventing it.
- Results travel to where you work: Slack, Notion, Zapier, HubSpot, Attio, Affinity, and as a file to Google Drive, Dropbox or Yandex Disk. Pushing notes and pushing action items are switched on separately.
- Open promises with an approaching date are raised by the agent — with the text of the promise and the meeting it came from. How that works from the agent's side is described in the agent that takes action.
The limits, honestly. The system does not assign the owner: it distinguishes ownerless work from a specific person's promise, and a human puts the name on the card. It proposes dates from what was said out loud — «by Friday» becomes the nearest Friday, which is not always right. Tasks are never closed or deleted automatically under any settings. The free plan gives 60 minutes of recognition per month; meeting analysis, the board and search across past conversations are on paid plans.
If what you need after a meeting is a document rather than cards, that is the adjacent problem, covered separately: from a recording to minutes that work. The wider picture of the category is in AI meeting assistant.
Frequently asked questions
How does AI find tasks after a meeting?
It breaks the transcript into objects: assignments, promises, decisions, questions. First over fragments of the conversation, then over the whole meeting — because a fragment does not know that twenty lines later the agreement is cancelled. Every found item is attached to a specific line; what cannot be attached is not created.
Will AI determine the owner and the deadline?
It will propose them, not guarantee them. «Sergey will look at the analytics by Friday» is parsed confidently; «let us update the onboarding» stays a task with no owner — and that is the right answer rather than a shortcoming. Names and dates in important items are checked by a human: thirty seconds per meeting.
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. Only creating your own tasks may apply automatically, and only if you enabled it. Calendar events, deletions and task closures stay manual always.
How is this different from an ordinary transcript?
A transcript answers «what was said», analysis answers «what we do». In a transcript an idea, a proposal and a promise look the same — they are all just sentences. Analysis tells them apart, which is why other people's thinking out loud does not end up in your task list.
What happens to promises made to a client?
A promise differs from an internal task in having an addressee, and that is stored along with the text. For sales and agencies it is half the point: «we will send the quote by Wednesday» without who it was promised to is a useless line.
What if nothing was decided in the meeting?
Then the list will be empty, and that is an honest result. Tasks cannot be extracted from a conversation without decisions — they are not in it. An empty analysis after an hour-long meeting is itself a useful signal: perhaps the meeting should have been an email.
What we do next
A good meeting ends not with «thanks everyone» but with a clear answer to the question: what do we do next? A list of four lines with names and dates is worth more than an hour of discussion — because an hour of discussion without that list is worth nothing.
You can check this on your next call rather than on an archive: create an account — 60 minutes a month are free — turn recording on and look at the list ten minutes after the meeting. A useful test: first write down from memory what has to be done, and only then open the analysis. The difference between the two lists is what you normally lose.