Meeting notes with AI: decisions, tasks and conditions
Turn a meeting record into clear notes with decisions. A worked example and prompt help preserve conditions, ownership and deadlines without guessing agreements.
Article language: English

In this article
“We could open the exhibition in June” does not mean “we are opening the exhibition in June”. That distinction can disappear from meeting notes when a neat summary is the goal. Prepare a record that also shows conditions and decisions still to be made.
The exchange below is fictional. This is an editorial exercise, with no recording of real participants:
Alice: We could open the exhibition on 12 June if we receive approval to borrow the exhibits by the end of May. Bartek: I confirm that I will check the transport terms by 6 May. Celina: I suggest printing 300 posters. Alice: We will decide how many posters after checking the budget. Let's agree now: Bartek will gather the transport terms and I will check the approval. We have not set a deadline for checking it.
Do not ask for “a five-point summary” yet. First ask the model to distinguish types of statement. A five-point limit should not encourage it to hide a condition in an imaginary sixth dimension.
A prompt that preserves uncertainty
Prepare working notes based on the fictional transcript.
Separate confirmed tasks, decisions, proposals, conditions and questions that need resolution.
For a task, include an owner, action and deadline only when the material supports them. Mark a missing deadline as “to be agreed”.
Quote a short supporting passage for each item.
Do not infer agreement to a proposal from silence.
Finish with two questions the meeting chair should clarify.
A correct example includes Bartek's task with a 6 May deadline and Alice's task without a deadline. The 12 June date remains conditional. Three hundred posters is a proposal, not an order to the printer. Two useful questions concern when to check the approval and how to determine the poster budget.
In this exercise you can check every word because you have all the material. With a real conversation, keep a way to refer to a specific passage in the recording or notes. A fluent account alone does not establish accuracy.
A transcript is not a record of decisions
A transcript records what was said, usually from audio. It may contain mistakes in names, amounts and speaker identification. OpenAI's documentation on files and audio recommends checking important details against the recording and points out that speaker identification can be unreliable.
If two product names sound alike, add their correct spellings as separate information to check. Do not silently correct an uncertain amount. Mark the passage for listening. “Fifteen” and “fifty” can create a very ambitious budget.
In Microsoft Teams, access to a transcript depends in part on roles and meeting settings, as explained in the official instructions for starting and downloading transcription. Before inviting an assistant or recording, check your organisation's rules and whether participants have been informed. A handwritten note using fictional data is enough for this exercise.
Test how the model handles a date suggested by someone else. Add this to the fictional conversation: “Celina: Could Alice check the approval by 4 May? Alice: I cannot confirm a deadline yet.” A new date has been mentioned, but it has not become Alice's commitment.
Update the exhibition notes after adding two statements.
Celina: “Could Alice check the approval by 4 May?”
Alice: “I cannot confirm a deadline yet.”
Show only items that need changing or adding.
Separate a proposed deadline from an accepted one.
Keep Alice's task without a confirmed date.
Quote the sentence on which your assessment rests.
A useful result adds Celina's suggestion but keeps Alice's deadline as “to be agreed”. It does not move 4 May from proposals to commitments. Give another person only the notes and ask when they think Alice must complete the task. If they read it as a firm date, the label is still unclear.
This is a practical form of checking AI answers against their sources (in Polish). You verify a specific statement rather than judging an entire summary with a single “looks good”.
Separate assigning a task from asking for confirmation
Show the working result to the meeting chair. Leave an empty field or a “to be agreed” label where a task's owner is unclear. The model should not pick an owner based on who spoke most.
You can draft a message: “The record mentions a conditional opening date of 12 June. Please confirm when we will check the approval to borrow the exhibits.” That asks for missing information. “I confirm the opening on 12 June” would create a new agreement.
Before sending a message, apply the guidance on prompts for emails (in Polish). Recipients should see which items require a reply and which are recorded for information. Do not automatically send the full transcript to everyone in your contacts.
Another difficulty arises after a participant corrects something. Add a fictional message after the meeting: “Bartek: I meant gathering the terms by 8 May, not 6 May. Please correct the deadline.” Do not erase the record of the change if the team has already worked from the first version.
Propose an update to the notes after Bartek's message.
Original record: gather transport terms by 6 May.
Bartek's correction: by 8 May, not 6 May.
Show the old entry, proposed new entry and source of the correction.
Mark the change as requiring confirmation by the chair.
Draft one short sentence informing the team of the proposed correction.
Do not change other dates or the conditional exhibition date.
Check that only the date of Bartek's task has changed. A later message from one person may need to be reconciled with others' schedules. In the exercise you propose an update and the chair finalises the agreement. Do not ask the model to decide who has organisational authority to move a deadline.
If you practise such tasks as part of learning, keep both versions and the reason for the change. The guide to choosing between an individual course and video (in Polish) shows how to use similar material to assess your independence. Being able to defend one correction says more about progress than the number of notes generated.
A useful check after the meeting
Check every name, amount and date that affects action. Compare tasks with the quoted passages. Then read the open-issues list separately: did the model treat an issue as unimportant simply because it was unresolved?
As a test, remove Bartek's deadline from the fictional record and repeat the task. The new version should show that the deadline is missing. If the model keeps 6 May, return to the input and check that it is using only the current material. Changing one detail this way tests whether the notes follow the source.
Store notes according to your organisation's access rules. We discuss limiting the audience in working safely with data in AI (in Polish). If you want to transfer tasks to a project system, use human approval of actions (in Polish) before the first suggestion becomes an official commitment.
Record the version and what was corrected when notes are amended. Someone returning to the agreement should be able to distinguish the original record, a proposed change and its final confirmation. If you share only an excerpt, include the condition that limits its meaning too; without it, the exhibition date can sound like a decision again.
In the AI Course (in Polish), we work through an example and discuss the result (in Polish). We develop this practice at Aeneia. At your next meeting, reserve the last two minutes to read decisions and conditions to the participants. It is a chance to correct the record while everyone still remembers what “let's try” meant.
*Sources checked on 30 October 2026. The conversation and sample notes are an original fictional demonstration.*







