The decision is made. Nobody knows by whom and by when.
The meetings themselves are fine. People come together, discuss, and decide. What breaks is everything after the last sentence. The decisions live in people’s heads and in three different private notes that do not match each other.
Tasks are handed out in the room, but not in writing. No named owner, no due date, no place where they would be visible the next morning. Two weeks later the same topic is back on the agenda, and the first fifteen minutes go into reconstructing what had long been agreed. The same topics get discussed three times before anything moves.
Whoever misses the meeting loses the thread completely. There are no minutes to read, so they keep working on outdated assumptions. The real damage is not the missing document. It is the decisions that were genuinely made and then quietly never executed.
Start with one meeting series, define what a task is, then automate.
Meeting automation fails when it is rolled out across every calendar entry at once. And the teams’ conditions become our specification: recording only with consent, opt-in per meeting, and sensitive rounds permanently excluded. One managing director put it like this: we do not want everything captured, we only want to know what was decided.
- Pick one recurring series: A project stand-up or a weekly leadership round. Recurring matters, so the team judges the AI employee over several cycles and not on one lucky session.
- Clarify the confidentiality rules first: Who consents, how a passage is marked confidential, which meeting types are permanently excluded, and how long minutes are stored. This conversation belongs at the beginning, not in a later data-protection review.
- Define what counts as a task: A task without an owner and a due date is a wish. We agree the mandatory fields with the team so the AI employee can flag what is incomplete instead of filing vague intentions.
- Run a few cycles in parallel: The team keeps writing minutes as usual while the AI employee drafts its version. Comparing two or three sessions shows exactly what is still missing and builds the trust needed to drop the manual version.
80 percent is done by the AI employee. The human decides at two points.
The AI employee joins the Microsoft Teams appointment as an announced participant, or it processes the notes someone writes anyway. From that input it pulls exactly four things: decisions, phrased as decisions and not as a discussion summary. Tasks with owner and due date. Due dates that were named in the conversation. And open points that were raised but not resolved.
The human stays in the loop at two points. The meeting lead releases the minutes, which open as an editable draft and go out through their own Outlook account. And every owner confirms their task with one click before it is written into the project tool. After that the AI employee reminds owners before the due date and proposes the next agenda from the open points. This is exactly why the same topic stops being discussed three times.
Recurring project stand-up, 45 minutes, 6 participants. Status of an ongoing rollout, two escalations, and a budget question at the end.
One field is flagged for review: the minutes only go out once the meeting lead has released them. The budget question gets no guessed owner, it moves to the next agenda.
That was before, this is today.
This is how meeting follow-up ran before, and this is how it runs today with the AI employee. Conservatively calculated, that saves about 90 percent of the time spent on minutes per meeting.
| Before | Today | |
|---|---|---|
| Writing the minutes | 20 to 30 minutes by hand | 2 to 3 minutes releasing a draft |
| Decisions | In heads and private notes | As a named decision in the minutes |
| Tasks | Handed out verbally, often without name and date | Owner and due date mandatory, confirmed by the owner |
| Repeated topics | The same topic three times on the agenda | Open points sit on the next agenda |
| Absent participants | Keep working on old assumptions | Read the released minutes, see their tasks |
* Savings conservatively calculated: writing and sending the minutes took 20 to 30 minutes per meeting before, today it is 2 to 3 minutes to release the draft. At 10 meetings per week that adds up to several hours. For context: leadership roles spend 15 to 20 hours a week in meetings, and the follow-up work sits on top of that. Not counted is the larger effect that decisions get chased at all.
How do you build an AI employee like this, technically?
The knowledge from this kind of project to take away, whether you build with us or on your own:
Listening: connecting to Microsoft Teams
The AI employee pulls the transcript through the Microsoft Graph API out of Microsoft Teams, directly from your Microsoft 365 tenant. There is no new meeting platform. Where no audio is wanted at all, the written notes of one participant are enough as input, and the structure behind it stays the same.
Extraction: separating decisions from tasks
From the transcript, a language model like GPT from OpenAI or Claude from Anthropic pulls four field types: decision, task, due date, open point. The hard part is separating what was decided from what was merely discussed. Exactly that distinction is the value of the minutes.
Never invent an owner
If a name and a date are named in the conversation, a task with owner and due date is created. If either is missing, an open point goes to the meeting lead. Every extracted task shows the passage in the transcript it came from. A task on a guessed name is worse than no task at all.
Task sync: one list, not two
Confirmed tasks are written through the API into the board the team already works in, for example Microsoft Planner or Jira, with owner, due date, and a link back to the minutes. A parallel task list that an AI employee creates itself stops being read within a month.
UX: the interface decides adoption
The draft minutes sit ready for release and stay editable until they are sent. Owners confirm their task with one click, due date included. And the open points show up as an agenda proposal for the next session. On top of that sits the opt-in rule set, where the team decides itself which meetings are recorded and which never are.
What does it cost in comparison?
Superkind charges per use case. The price grows with the number of meeting series, not with headcount. Here is the honest comparison:
| Handwritten notes | Transcription tool | Superkind AI employee | |
|---|---|---|---|
| Cost | No license, but 20 to 30 minutes of working time per meeting | 10 to 30 € per user per month | Price per use case, a fraction of a full-time position |
| What is included | Minutes by hand, varying by person | Only the full-text transcript | Minutes, tasks with owners, filing in the project tool, reminders |
| Scales with | More participant time | Number of users | Number of meeting series, without new positions |
| Exceptions | Fall through the cracks | Stay buried in the text | Open points move flagged to the next agenda |
| Rollout | Nobody feels responsible | Creating an account | 2 to 3 weeks to the first productive version |
A transcript is not a follow-up anyone is chasing. Transcription tools deliver a complete text, but no owner, no due date, and no reminder. The honest comparison is the full cost of manual follow-up: the time on minutes, the verbal chasing, and the decisions that were quietly never executed.
What we learned from projects like this.
The reflex with meeting tools is to capture everything: full transcripts, searchable archives, every word preserved. In our experience that is exactly why such projects get rejected. People speak differently when everything is captured, and a team that speaks carefully has worse meetings, not better documentation. The AI employees that get accepted are the ones that capture less: four fields, nothing else.
The second thing: no AI employee fixes a meeting culture. If a round ends without a decision, the honest minutes say that nothing was decided. Some teams find that uncomfortable enough to blame the tool. That should be said out loud before building: this makes commitment visible, and visibility only helps teams that want it.
What it is not suited for: Confidential bodies like HR conversations and works-council meetings should stay unrecorded by design. One-on-ones and feedback rounds depend on nothing being written down. And if meetings regularly end without anyone deciding, the AI employee will reliably record that nothing was decided. That is a leadership topic.
