You Didn't Get a Cost Reduction — You Got an Army

your meetings aren't the problem. your team size is.

i've been saying this for a while and people push back. "dom, we're busy. ai note-takers help." no. they don't. they're barnacles on a broken ship. the ship is your team structure - and AI just made the problem five times more expensive.

let me explain what's actually happening.

the math nobody is doing

the number of communication pathways between people in a group isn't linear. it's combinatorial.

5 people = 10 pathways. manageable. one person can hold the full map in their head.

10 people = 45 pathways.

20 people = 190 pathways.

Robin Dunbar figured this out in 1992 from primate neocortex research. the human brain has hard limits on relationship complexity - about 5 for deep coordination, 15 with trust, 50 for meaningful working relationships.

the military confirmed it empirically. a US infantry fire team is four people plus a leader. the layers above track Dunbar's hierarchy almost exactly. Jeff Bezos landed on the same number from a completely different direction - the two-pizza team. Fred Brooks got there in 1975 through software engineering: adding people to a late project made it later, not earlier. i still see executives who don't believe this. they're wrong.

three completely separate disciplines - evolutionary psychology, military strategy, software engineering - all converged on the same answer. the human brain can sustain deep, high-context coordination with about five people.

AI didn't change that number. it didn't rewire our brains.

what it changed is the cost of getting the number wrong.

what AI actually did to your team

before AI, a five-person team produced X output. adding a sixth person gave you more capacity but with diminishing returns. Toby Lutke at Shopify calls it a 10x loss of productivity with each addition beyond five. that coordination overhead has always existed.

after AI, the same five-person team produces 5 to 10 times more than before.

you can see this in the revenue-per-employee numbers of AI-native companies. Lovable, hitting hundreds of millions ARR with 45 people. Midjourney, same story. ElevenLabs, same story. the SaaS benchmark for revenue per employee has historically sat below $500k. AI-native companies are running 5 to 10 times that.

so here's what changes: if each person on your five-person team is now producing $2-3M a year in value, the coordination cost of a sixth person is no longer a minor tax. it's a catastrophe.

that sixth person doesn't just need to be good. they need to justify their coordination cost against a baseline where every existing member generates output that previously required an entire department.

when each person produced $250k a year, the coordination cost of person six was manageable.

at $2M per person, it's measured in millions of lost productivity.

the bar didn't rise a little. it rose by an order of magnitude.

why your meetings are killing you

every meeting exists because someone decided coordination was worth the cost.

when your per-person output was $250k, it often was worth the cost. at $2M per person, most of those meetings are now net negative. they destroy value at a rate that scales directly with how productive your people are.

studies put the average knowledge worker at 12 hours in meetings per week. people managers, 16 hours. execs, 23 hours. and those numbers keep going up, not down.

AI didn't fix this. AI made it worse - because it increased output per person without touching the coordination overhead. you're now burning $2M-per-person time in alignment sessions that produce more alignment sessions.

the standup that could be a Slack message. the cross-functional sync where eight people attend and two people talk. the alignment meeting that schedules the next alignment meeting.

you don't have a meetings problem. you have a team size problem that's generating a meetings problem.

volume is free. correctness is scarce.

here's what keeps getting missed in every AI and teams conversation i hear: people obsess over volume. more code, more content, faster.

that framing leads to disastrously wrong organizational decisions.

a Harvard Business School field experiment published in 2025 tested 776 professionals at Procter & Gamble on real innovation challenges. teams using AI were three times more likely to produce ideas in the top 10% of quality. not three times more output. three times more likely to be right at the highest level.

AI made volume free. what's scarce is correctness - whether the thing you shipped is architecturally sound, strategically coherent, right for the customer, free of the subtle errors that look fine in a demo and compound into real failures in production.

a team of five optimizes for correctness. every person's AI output passes through at least one other brain that shares enough context to catch meaningful errors.

a team of 20 is optimized for volume. in a world where AI makes volume free, optimizing for volume is optimizing for the wrong thing.

this is why big teams can feel productive - lots of Jira tickets, lots of activity - while still shipping things that don't quite work, that need rework, that spawn follow-up projects to fix the problems the last project created.

volume masquerades as progress. correctness is progress.

two archetypes that actually work

i've started thinking about two structural units - and they fit different missions.

Scouts: one person, full AI toolkit, defined mission. zero coordination overhead. the constraint is one person's judgment - which means fast for exploration, but no error correction. great for: is this technology viable? is this market real? can we prototype this in a week?

Peter Steinberger demonstrated this at its extreme. in roughly 60 days, running 4-10 coding agents simultaneously in Codex, he built OpenClaw - in a language he'd never used. he directed agents at the architectural level while they handled execution. one person, 20 years of judgment, a swarm of agents. the world's most valuable companies were desperate to acquire it.

but scouts have limits. OpenClaw shipped with holes. solo doesn't work when correctness requires multiple perspectives or when sustained production is the goal.

Strike Teams: five people, AI-augmented, executing where correctness matters. a team of five can cover product, engineering, design, data, and domain expertise. that's the minimum surface area for a complete decision. below five you have blind spots. above five you have silos.

in a team of five, there is nowhere to hide - which is exactly what you want.

most organizations right now have neither of these. they have oversized teams that are too slow for exploration and too diluted for precision execution. burning their best people on coordination overhead and wondering why AI isn't working.

the reframe everybody misses

here's what drives me crazy about the AI and team size conversation.

everyone frames it as a cost story. "we can do the same work with fewer people." that's the headline. that's the strategy deck.

it's a staggering failure of imagination.

if you have 500 people and each just got 5 to 10 times more capable - the correct response is not "i can run this company with 50 people."

the correct response is: "i have the capacity of 2,500 to 5,000 people. what was i previously unable to do?"

your 500-person company just acquired the productive capacity of a 3,000-person company without hiring anyone, without raising capital, without building new offices.

you did not get a cost reduction. you got an army.

the question is whether you have the strategic vision to deploy it - or whether you're going to run a fleet of aircraft carriers on the same fishing route your trawler used to run.

Lovable didn't use AI to shrink. they used a small team to think really, really big. Midjourney didn't look at 100 employees and say "we need to focus on our niche." they went after the entirety of visual creation.

these companies didn't use AI to cut. they used small teams to expand what was possible.

how to actually move

Toby Lutke at Shopify required every team to prototype with AI before beginning a real build. every project, every team. AI-first as a default, not an option. he made AI fluency part of performance reviews and required teams to demonstrate why AI could not do a task before requesting headcount.

the surface reading: "aggressive AI adoption push." the deeper reading: he built a systematic evaluation pipeline. every AI prototype generates a data point on what AI can and can't do in that domain. when the next model drops, they have a pre-built test harness.

and every forced prototype is a training rep for the skills strike teams need. the person who prototypes 10 times and fails 7 has built more specification skill than the person who attended 10 meetings on AI strategy.

you don't learn to direct agents by talking about it.

if you want to know who on your team is ready for the strike team model - give someone a real scout mission. a problem your company has been ignoring. full AI tooling, one week, clear objective, zero check-ins.

what you're testing for: can they define the problem without being handed a spec? do they know what right looks like at the architectural level, not the syntax level? do they default to action or permission?

the results won't match your current performance review rubric. some of your highest-rated people - great at running meetings, writing clear status updates, navigating org structures - may struggle. those are coordination skills. valuable in a big team. overhead in a strike team.

your most frustrating people - the ones who skip meetings, build things without asking, occasionally ship something brilliant that nobody requested - may be exactly who you need.

they've been fighting your organizational structure for years. the strike team is the structure they were built for.

the real question

the right question for any org right now isn't "how small can we get?"

it's: given that every five-person team now has the capacity of a 50-person department, how many teams do we need to pursue the mission we actually want - not the mission we settled for when headcount was expensive?

for some companies, the honest answer will be fewer people. the coordination roles and management layers that existed only because the org was too big to manage itself - those go away. that's real.

but for many companies, the right answer is: keep your people, restructure how they work together, and go after something bigger than what you're doing today.

the leaders who figure this out aren't talking about efficiency. they're talking about capability they never had before.

the ones who don't are going to spend a lot of time in AI meetings. with very good AI notes. optimizing a structure that's already obsolete.

are you using AI as a cost reduction - or as a force multiplier pointed at a bigger mission?

---

part of my "how i vibe" series on agentic engineering and team structure.

#howivibe #agenticengineering #softwareengineering #leadership