Every Monday we hold a leadership meeting with twenty managers around one table. If you price that hour by the salaries of the people in the room, it is the most expensive hour of the company’s week — and a year and a half ago I honestly admitted to myself that we were burning most of its value. People arrived with varying degrees of preparation, someone typed up the minutes late at night from memory, and decisions that ended up without an owner quietly dissolved by the next Monday, only to resurface a month later as “didn’t we already decide this?”
The only consolation was learning that we were not alone. Deloitte’s research on meeting culture cites a figure that should make the whole corporate world blush: roughly 40% of action items from meetings simply get lost — not executed poorly, but lost, because they were never captured with an owner and a deadline. The guidance on effective meetings has long converged on one cure: every decision gets exactly one owner and one deadline, and capture happens at the meeting, not from memory in the evening. Industry surveys from 2025 add that about three out of four regularly meeting teams have already adopted AI note-takers. The world, in other words, is moving exactly this way — the only question is how to make the note-taking part of management rather than a toy.
Since June, our Monday meeting has run entirely through our own AI loop. In two months the agents have processed more than four hundred meeting recordings — leadership meetings, client calls, sales series — and the pace now holds at about seventy recordings a week. Let me walk through what changed and why the result is not “a transcript from a neural network.”
What it looks like
Preparation starts before the meeting. Every participant assembles their package in a chat with the agent — items for decision, items “for information,” statuses of previous action items — and any employee in the company, even one not attending, can submit a question to the proposal board, where the agent checks it for duplicates. The facilitator needs a single command to get a pre-read — a consolidated agenda by workstream. Over two months facilitators have assembled 32 such pre-reads and participants have saved 172 preparation packages, and the meeting now starts with decisions instead of the ritual “tell us what you’ve been up to.”
The meeting itself runs as usual; the agent does not interfere in any way. When everyone leaves, the facilitator simply drops the audio recording into the chat, Whisper transcribes it locally — the recording never leaves our perimeter — and a few minutes later the chat shows the main thing: a draft of the minutes with summaries by workstream, decisions and action items. This draft used to cost one to two hours of manual work late in the evening; now it costs minutes of review. The “Minutes” document appears in the system only after the facilitator’s explicit “yes”: nothing is published without a human, and that is not a precaution for the early days but a permanent rule of the loop.
An action item = one owner + one deadline
I consider this rule the most important one in the entire loop, and it was gratifying to discover that the global playbooks preach exactly the same thing — “one owner plus one deadline,” or you get those forty percent of losses. A verbal agreement without an owner is a wish, not a decision, and the agent is raised in the same logic: it records as an action item only what has an owner and a deadline, while ownerless phrasings are honestly set aside in a separate list of open questions rather than disguised as decisions.
The rule has a second half without which the first would not work: action item statuses automatically appear in the preparation for the next meeting. The question “what did we decide last time?” no longer requires archaeology through chats and someone’s memory — the thread between meetings is now pulled by the system, not by the most conscientious participant.
Why this is not a transcript
Turning audio into text is the easiest part of the task, which is exactly why solutions of the “just transcribe our meeting” class stumble so uniformly over everything that begins after it.
The first obstacle is names. Transcription confidently writes distorted versions of real names and of the d8n brand, and if you leave that as is, the minutes turn into a puzzle. Our agent matches every owner against the corporate registry by phonetic similarity and writes the canonical full name; when there is no unambiguous match, it flags “needs confirmation” instead of guessing.
The second is structure. The text of the minutes, the action item table and the full transcript live in three different places with no duplicates, and the document itself exists in the standard document workflow rather than as a file on the facilitator’s drive — a drive that will one day go on vacation along with its owner.
And the third is control. A two-phase principle applies to every action: the draft is always shown to a human, and the document and the decision register are created only upon explicit confirmation.
What it delivered
The time arithmetic is simple: one to two facilitator hours per protocol turned into minutes, which on the weekly series alone amounts to roughly 50–100 hours of management time per year. An honest caveat — this is an expert estimate; we did not run a time study.
But the main effect, as often happens, was not in the hours, and I have written many times that saved hours are a weak metric. The real change is in execution discipline: an action item with an owner, a deadline and an automatic status at the next meeting is a different level of manageability at zero additional load on participants. Decisions stopped dissolving — and by feel, that changes the meeting more than any saved minutes.
A pleasant bonus: the loop scales without development. A new recurring meeting — a department stand-up, a product series — connects with a line of configuration. Our sales series launched exactly that way.
Key takeaways
- The most expensive hour of the company deserves an engineering approach no less than a production process does.
- An action item without an owner and a deadline is not a decision. Let the system refuse to call it one.
- Transcription is 20% of the task. The other 80% are names, structure, statuses between meetings and human control.
- Start with minutes from audio: the effect shows on day one, and preparation and the decision register connect afterwards.
All of this is from our own operations data, May–July 2026. You can see the minutes scenario with your own eyes in the live demo.
