TL;DR AI made individuals 5-10x more productive. It did not reduce the coordination cost between them. That mismatch means every person you add above ~5 on a team now costs millions in lost output. Most orgs are still staffed for 2019. The ones that restructure around small teams, correctness over volume, expanded ambition will compound faster than competitors realize what happened.
I recently started a new CTO role and did what I always do first: audit where my time goes. Meetings. The standard take is that we have a "meetings problem" with a tooling solution: AI notetakers and calendar audits.
I think that diagnosis is wrong. We don't have a meetings problem. We have a team size problem that manifests as meetings. Every AI productivity tool we adopt without fixing the underlying structure just makes the tax bill bigger.
The formula everyone knows and nobody applies
Communication channels scale as n(n-1)/2. Five people, 10 links. Ten people, 45. Twenty people, 190. Brooks wrote about this in 1975. Almost nobody uses it to make staffing decisions.
Here's what changed. The value flowing through each of those links went up by an order of magnitude. Cursor hit a billion-dollar run rate with around 300 employees. Revenue per employee is a noisy metric, but even after discounting, the productivity gap between AI-fluent teams and everyone else is real, it's widening, it changes the math on coordination cost.
When your five-person team generates $300K per head, adding a sixth carries a tolerable coordination tax. Maybe $50K in drag. When that team generates $2M per head, a sixth member who introduces even modest friction isn't costing you fifty grand. They're costing you millions.
Tobi Lütke at Shopify called it a "10x loss of productivity" for each addition beyond optimal.
Three fields walked into the same answer
Robin Dunbar's primate neocortex research (1992) found that human social groups cluster at specific sizes, with about 5 for the tight inner circle. The military figured this out independently. A US infantry fire team is four soldiers, team leader included. The layers above (squads, platoons, companies) track Dunbar's hierarchy almost perfectly.
Fred Brooks got there from software engineering. His law applies to late projects specifically, but the mechanism generalizes: coordination overhead grows faster than added capacity. AI made it worse because it amplified the output being taxed.
Four independent paths (evolutionary psych, military doctrine, software engineering, corporate operations) to the same conclusion: humans sustain deep coordination with about five people. AI didn't change that number. It changed what you lose when you get it wrong.
Volume is free. Correctness is everything.
Everyone's measuring throughput. More PRs merged, more content shipped, more Jira tickets closed. Volume is not the scarce resource anymore. AI made volume cheap. Optimizing for cheap things is a strategic error.
A 2025 Harvard Business School field experiment ("The Cybernetic Teammate") studied 776 professionals at Procter & Gamble on actual innovation challenges. AI-augmented teams were about three times more likely to produce ideas in the top 10% of quality.
That's the mechanism. Five good generalists using AI can cover more surface area than ten specialists in narrow lanes. But the mutual verification layer only works when everyone shares enough context to catch each other's mistakes. In a five-person team, you can hold the full model in your head. At twenty, you can't. So you hold meetings to synchronize, which generates more output to verify, which generates more meetings.
Wes McKinney, the pandas creator, nailed this in "The Mythical Agent-Month". He describes a "brownfield barrier": agent-generated codebases hit around 100K lines then the agents start choking on the bloat they themselves created. Design taste, product scoping, holding the conceptual model in your head are now the binding constraints. Judgment. McKinney moved to Go for new projects because its fast test cycles work better for agent iteration.
Two modes: exploring and building
Scouts are 1-3 people doing exploration. Is this viable? Is the market real? Can we prototype it? Near-zero coordination overhead. Speed is everything. The constraint is individual judgment, full stop.
Peter Steinberger's OpenClaw is the poster child. One developer directing a swarm of AI coding agents at the architectural level. Incredible velocity. Also shipped with significant security holes that Cisco's researchers documented, because solo work doesn't have a verification layer. Steinberger joined OpenAI in February 2026.
Strike teams are 3-7 people executing where correctness matters. Every piece of AI output passes through at least one other human who shares enough context to catch real mistakes. Below five you get blind spots. Above seven you get silos plus meetings about meetings.
Most orgs have neither. They have twelve-person teams that are too slow to explore and too diluted to ship with precision.
The ambition problem nobody talks about
The entire conversation around AI and headcount is framed as efficiency. That's the least interesting thing you can do with a force multiplier.
If you have 500 people who are each 5-10x more capable than two years ago, you didn't get a cost reduction. You got the productive capacity of 2,500 to 5,000 people without hiring anyone. The question isn't "how lean can we get?" It's "what was previously impossible that's now on the table?"
Lütke understood this. His April 2025 memo required every Shopify team to prototype with AI before starting a real build. Teams have to prove AI can't do a task before requesting headcount. Every forced prototype generates a data point, so when the next model generation lands, Shopify already has the test harness ready.
The companies that will matter in five years aren't cutting people to protect current margins. They're restructuring into strike teams pointed at missions that were unthinkable when each person produced $300K a year. The constraint stopped being capacity. It's imagination now.
Where I think this argument is wrong, or at least incomplete
The biggest hole: small teams disrupt, but large teams develop. Wu, Wang and Evans analyzed 65 million papers, patents, software projects in a 2019 Nature study. Small teams generate novel ideas. Large teams consolidate and integrate. You need both. The strike team isn't your answer for a platform doing millions of transactions daily at five-nines uptime. It might be your answer for the feature teams building on top of it.
Then there's the Spotify problem. Their squad model was the most famous implementation of small autonomous teams. Spotify itself quietly abandoned it. Jeremiah Lee documented what went wrong: squads optimized locally at the expense of the whole, knowledge got trapped, dependencies created deadlocks nobody owned. Making teams smaller doesn't eliminate coordination. It moves the problem up one level.
An NBER study in February 2026 surveying nearly 6,000 executives found over 80% reporting no AI productivity impact yet. Mary Daly at the San Francisco Fed compared it to electrification: factories had electric motors for years before anyone redesigned the floor around them. Executives who cut headcount based on AI-native startup data before their own teams have crossed the fluency threshold will hollow out institutional knowledge for a productivity dividend that never arrives.
And five is not a magic number. A 2021 replication of Dunbar's method got 95% confidence intervals ranging from 2 to 520 people. The directional argument holds. The specific number is a heuristic, not a constant.
So what do you actually do
Measure correctness instead of velocity. Count post-launch defects, rework cycles, integration failures. If the defect rate is climbing alongside the output rate, you're shipping faster toward the wrong destination.
Run scout missions to map your talent. Hand someone a problem, full AI tooling, a week, no committee oversight. Watch the process. You're looking for people who define problems without a spec. The ones who run beautiful meetings have coordination skills. Valuable in a big team, overhead in a strike team. Meanwhile your most frustrating employees, the ones who ignore process and build things nobody requested, might be exactly who you need.
Build the ambition pipeline first. What would you build if every five-person team had the capacity of a 50-person department? If you can't answer that, you don't have a productivity problem. You have an imagination problem.
Redeploy, don't fire. The team that automates reconciliation doesn't get cut. Their capacity moves to the product problem you could never staff before. But that only works if the ambition pipeline already exists.
Most orgs will keep running twelve-person teams, burning half their week in meetings, deploying AI notetakers to summarize the waste. A few will restructure. Small teams. Big missions. Correctness over volume. Those are the ones I'm betting on.