I watched a team of 42 install an AI notetaker in January and, by March, had bought a second AI scheduling tool to go with it. Their calendar looked identical to how it looked the month before. The tools worked beautifully. Not a single meeting got smaller, shorter, or cancellable. This is the gap I want to close in this post, because it is the one every remote team hits in 2026: the category matured, the behavior did not.
McKinsey State of AI 2026, published in August 2026 from a survey of more than 1,700 executives across 97 markets, puts a surprising shape on it. Eighty-three percent of respondents said AI’s productivity and focus value came from enabling async-first work. Fifty percent said AI helped them make better decisions. Buying tools moved nothing on its own. What moves is the default state of a team’s collaboration. If the default is still synchronous, an AI notetaker is just capturing your waste at higher resolution.
Do the arithmetic before you buy anything. Six recurring meetings a week with thirty attendees each is 180 attendee-hours per week, about 9,360 hours a year, roughly five full-time equivalents of paid time going into rooms that never end in a written decision. Gallup State of the Global Workplace 2026, based on 263,810 respondents, finds global engagement at 20 percent, with disengagement costing roughly 10 trillion US dollars globally and 322 billion in the United States. Low-quality synchronous time is a line item on that number, and it is the cheapest one to cut.
Step one: classify before you schedule
Every recurring meeting falls into exactly one of three classes, and only two of them belong on a live calendar.
Announce
Someone has information, everyone else needs it, and no one is expected to respond. Roadmap updates, incident summaries, status rollups, hiring news, budget results. These are broadcasts. A broadcast that requires a room is a bug, not a design choice. An announce becomes a written note posted to the channel people already check, and the meeting dies on its own.
Discuss
Genuine uncertainty is on the table and the answer lives in the interaction itself. Trade-off exploration, ambiguous requirements, cross-team dependency knots. These earn synchronous time because compression matters and reading a paragraph of hedging is slower than talking through it for nine minutes.
Decide
A choice is pending, someone is blocked behind it, and the decision maker is in the room. This is the most expensive meeting to run and the most important one not to cancel. The failure mode here is not that it happened. It is that it happened without a written decision record that you can quote to someone who was not invited.
I classify every recurring meeting on my own team once a quarter. The share of recurring meetings that survive classification has been dropping every quarter for two years, and nobody notices, because nothing gets louder when an announce disappears.
Step two: run the 24-hour async protocol
For the discuss and decide meetings that survive, the sequence is fixed and non-negotiable.
Day minus one
Post the pre-read and the decision document at least 24 hours ahead of the invite. The pre-read should fit in a single screen. The decision document should name the options, the stakes, the recommended option, and the decision owner. The decision owner is the person who owns the outcome, not the person who wrote the document, and the two are often different people.
The meeting itself
Open by reading the decision line aloud and asking whether anyone has a material objection. If the pre-read was read, you will know in under two minutes. Timebox the discussion to the decision, not to the background, and assign one person to the note even when the AI is running, because the AI transcribes everything including the parts that were explicitly discarded.
Day plus one
Any AI-generated note must be confirmed or corrected by a named owner within 24 hours. The rule is simple: an unconfirmed note is not the record. This is the only part of the protocol people push back on, and it is the part that makes the rest work.
Step three: cap attendance by role
The attendance rule I use is a cap, not a preference: the accountable owner, the decision maker, and the two people with unique context. That is four people. Everyone else gets the notes and the decision record, and is welcome to object in writing within 24 hours.
This single rule has cut my team’s total meeting load by more than half. The uncomfortable part is that most attendees do not know they were attending to feel informed rather than to contribute. Naming the role they were filling, and then removing the room, is the fastest culture change available.
| Meeting class | Default channel | Sync time allowed | AI notes needed | Success measure |
|---|---|---|---|---|
| Announce | Written note in the team channel | None | No | Read receipt within 48 hours |
| Discuss | Thread or doc comments, then a timeboxed call | 30 minutes maximum | Yes, confirmed within 24 hours | Question resolved or escalated |
| Decide | Decision doc, then a call with the owner only | 20 minutes maximum | Yes, confirmed within 24 hours | Written decision with named owner |
Two of those rows keep recurring meetings alive. The announce row is the one that keeps killing them, and that is the point.
Step four: measure one quarter of behavior
Culture change without measurement is a mood. I track three curves weekly and review them at the end of each quarter.
- Meeting count per person per week. This should fall first and fastest.
- Synchronous hours per person per week. This follows, and it is the number that frees real capacity.
- Action-item completion rate. This is the one that must not fall. A team that cut meetings and cut execution just moved the failure from calendars to deliverables.
When the third curve drops, do not read it as culture succeeding. Read it as your async protocol failing, and go back to the 24-hour rule. The pattern I want to see is meetings falling and completion rising or holding flat. Anything else is noise.
Where this stops working
The framework breaks in two specific situations, and I want to name them because pretending otherwise is how teams over-apply it. First, genuine trust repair, where a team is rebuilding after a breakdown, still needs synchronous time, and cutting meetings in that window makes it worse. Second, teams under ten people with one customer in front of them cannot afford the abstraction, and a short standup beats a decision document until the team has enough surface area to hurt itself.
For everyone else, the arithmetic is unambiguous. Classify, move announces, cap attendance, measure for a quarter. The tools are not the change. The default state is.