The most powerful AI models on the planet share one counterintuitive habit: they refuse to use their whole brain for every question. DeepSeek V4, GPT-5, and Gemini 3 each hold more than a trillion parameters, yet answer most prompts by activating only a few billion of them. The technique is called mixture of experts, and in 2026 it quietly became the standard blueprint for frontier AI.

Apply the same idea to your calendar and you get mixture of experts for meetings: instead of inviting the whole team to every discussion, you route each decision to the two or three people who actually have context — and let everyone else stay heads-down. Most meetings do the opposite. The average knowledge worker now sits through 21.7 meetings a week, and Microsoft found that 57% of them are ad hoc calls with no invite or agenda. That is dense activation, and it is the most expensive way to run a team.

In this comparison, we'll put the dense meeting (invite everyone, just in case) next to the sparse one (route the decision to the experts) and score them on what actually matters: decision quality, speed, cost, and what happens to the people you leave off the invite.

How Frontier AI Stopped Wasting Compute

For years, AI models were dense: every parameter fired for every input. Doubling a model's knowledge meant doubling the cost of every answer. Mixture of experts broke that link. Instead of one monolithic network, an MoE model is split into many specialized sub-networks — the "experts" — sitting behind a traffic cop called a gating function. When a query arrives, the gate routes it to the handful of experts best suited to it and leaves the rest dormant.

The result is counterintuitive: a far bigger model that costs far less to run. DeepSeek's V4 Pro, previewed in April 2026, carries 1.6 trillion parameters but activates just 49 billion of them per query. As Hugging Face explains, that lets you scale a model dramatically on the same compute budget as a dense one. Specialization plus routing beats brute force.

That trade — more total expertise, less wasted activation — is exactly the one your meetings keep getting wrong.

Your Meetings Run in Dense Mode

A dense meeting is the organizational version of a dense model: when in doubt, activate everything. Need to make a call? Invite the whole team. It feels safe, and Harvard Business Review explains why it's the default — "it feels easier for the organizer to just invite an entire team or several, both to ensure that the right people will be there and to avoid offending anyone."

The bill comes due quietly. Asana's research finds knowledge workers lose 103 hours a year to unnecessary meetings and spend 60% of their time on "work about work" rather than the job they were hired for. Atlassian pegs the waste even higher, with poorly run teams spending 50% more time in unnecessary meetings than on priority work. One study put the price of pointless meetings at roughly $25,000 per employee per year — about 31% of every meeting dollar spent. We've broken down the full cost of meeting overload before, and the true price of a single meeting is almost always higher than it looks.

Dense activation isn't just slow. Like an over-parameterized model, it actively degrades the output. Which brings us to the sparse alternative.

Mixture of Experts for Meetings: The Sparse Alternative

Mixture of experts for meetings flips the default from invite to route. Before a meeting even exists, you ask the gating-function question: which two or three people have the context to move this decision forward right now? Those are your experts. They go in the room. Everyone else gets the output — the decision and the reasoning — without burning an hour to receive it.

This isn't "have fewer meetings" advice. It's a routing model. The meeting still happens; it just activates the right sparse subset of your team instead of the whole network. And like a well-tuned MoE model, the smaller activation usually produces a better answer, not a worse one.

Dense vs. Sparse Meetings, Compared

Here is how the two models stack up across the dimensions that decide whether a meeting was worth holding.

Decision quality

This is where sparse wins most clearly. Bain's research identified a Rule of Seven: every person added to a decision group beyond seven cuts decision effectiveness by about 10% — a group of 17 or more "rarely makes any decisions." Stanford's Robert Sutton reached a similar verdict, telling HBR that the most productive meetings hold five to eight people, past which conversation quality erodes. Dense meetings don't just cost more — they make worse calls. If team decision-making keeps stalling, the invite list is usually the first suspect.

Speed

Fewer, higher-context people reach decisions faster. There are fewer opinions to reconcile, less social hedging, and no one in the room waiting for permission to speak about a topic they barely touch. A pod of four experts can commit in the time it takes a room of fourteen to finish going around the table. Sparse activation is the quiet engine behind every team with high decision velocity.

Cost

Dense meetings multiply the most expensive resource you have: senior attention. Every extra head is a salaried hour spent listening instead of building — and, as the numbers above show, most of that listening changes nothing. Sparse routing treats attention the way MoE treats compute: spend it only where it moves the result.

What happens to everyone else

In an MoE model, the dormant experts aren't deleted. They're simply not activated for this query, and they still share in the model's overall knowledge. The team equivalent is an async output layer: the people you didn't route still receive the decision, the context, and the canvas — just not the live invite. This is where capturing the room matters. Coommit's contextual AI watches both the canvas and the conversation, so the two or three experts who met can hand everyone else the actual decision and its reasoning, not a vague "FYI" recap. Done well, this is the backbone of a healthy async work culture.

The failure mode

Sparse models have one classic bug: a lazy router sends everything to the same popular expert, who becomes a bottleneck. Teams have the identical failure. Route every decision to your one indispensable person and you've built a single point of failure, not a system. Good routing spreads load — grow more than one expert per domain so the gate always has somewhere else to send the query.

Who Should Be in a Meeting? Build a Gating Function for Your Invite List

You don't need software to route — you need a rule. Before sending the next invite, run each name through a simple gate:

Jeff Bezos's two-pizza rule is the same instinct in different words: if two pizzas can't feed the meeting, you've over-activated. The gate above just makes the cut feel like routing instead of exclusion — you weren't left out, your token simply wasn't sent to you. Protect the people you don't route, too: an uninterrupted afternoon of focus time is worth more than a seat in a meeting they'd spend half-listening to. Pair that with a few no-meeting days and the sparse model becomes the default, not the exception.

Conclusion

Frontier AI got cheaper and smarter the moment it stopped activating everything for everything. Your team can make the same jump. Mixture of experts for meetings isn't about banning meetings — it's about routing each one to the people who can actually move it, and giving everyone else the output instead of the invite. Start with your next recurring meeting: cut it to the experts, and send the rest the decision log.

The hard part has always been knowing who holds the context. That's the gating function worth building — and the reason contextual tools that see both the canvas and the conversation, like Coommit, are becoming the routers for how modern teams actually work.