your organisation chart is legacy code. stop operating like a dinosaur!
companies exist because talking was expensive. talking just became free.
in 1937, a 26-year-old economist asked a question so basic almost nobody had thought to ask it: if markets are so efficient, why do companies exist at all? why isn't everything just individuals contracting with individuals?
ronald coase's answer eventually won a nobel prize, and it fits in a sentence: because using the market is expensive. finding the right person, negotiating the deal, writing the contract, checking the work, doing it all again next week. those costs, transaction costs, made it cheaper to pull people inside one building where a manager just tells them what to do.
read that again. the company is not a law of nature. it's a workaround for expensive coordination.
firm size is a function, not a fact
once you see it, company size stops being a mystery. a firm grows until coordinating inside costs more than contracting outside. every technology that cut outside-coordination costs shrank the necessary firm. the telephone did it a little. the internet did it a lot: outsourcing, saas, apis, a solo developer renting amazon's data centers instead of employing a server team.
ai cuts both sides at once. that has never happened before.
inside the firm: most of what a company spends on coordination is humans moving information between humans. status meetings are synchronous blocking calls. middle management is middleware: it routes messages, buffers requests, translates between layers. eight layers of hierarchy is a caching strategy from an era when information lived in heads and moved at meeting speed. agents collapse all of it. context that took a department to hold now fits in a harness one person runs.
outside the firm: finding, vetting, contracting, verifying. the exact frictions coase named are precisely the tasks ai does well.
when both curves fall, the equilibrium size of the firm falls with them. not to zero. toward the irreducible core.
the evidence is already loud
the numbers coming out of ai-native companies don't look like efficiency gains. they look like a different species. lovable added $100m of revenue in a single month with 146 people: roughly $2.8m of revenue per employee, against a software median around $130k. a quarter of a recent yc batch shipped codebases that were 95% ai-written, in the fastest-growing batch in yc's history. the highest revenue-per-head companies ever built are being founded right now, and they are small on purpose.
none of these companies deleted their org chart. they never wrote one. that's the tell. the advantage isn't the model, everyone has the same models. the advantage is having no legacy structure to defend.
why "legacy code" is the right insult
legacy code doesn't crash the day it becomes legacy. it keeps running. everyone knows it's wrong, and nobody dares refactor it, because the people who understand why it's shaped that way have power precisely because it's shaped that way. so it quietly gets more expensive every quarter, the interest compounds, and the rewrite gets scarier the longer you wait.
that's a big company's operating structure in 2026. written for constraints that no longer exist: human-speed communication, information hoarded in heads, proximity as the price of trust. still running. increasingly expensive. defended by exactly the people a refactor would demote.
and the dinosaur part isn't an insult, it's a mechanism. dinosaurs didn't die because they were weak. they died because the climate they were optimized for stopped existing, and optimization is exactly what makes you fragile when it does.
coase isn't the whole story
firms also exist for things that aren't transaction costs: trust, culture, tacit knowledge, the ability to make decade-long bets, a legal body that can be sued and can therefore be trusted with liability. courts don't contract with a swarm of agents. big customers don't either. regulation, procurement, and enterprise sales still reward being big enough to sue.
so no, the firm doesn't disappear. it shrinks toward the parts that were never coordination: judgment, taste, responsibility, long-horizon conviction. the 10,000-person company was always a small core of judgment wrapped in a giant shell of coordination. ai eats the shell. the core stays human, and gets more valuable.
the squeeze lands hardest in the middle: companies big enough to carry the coordination shell, too small to have platform-scale advantages. very small and very large both have a future. the middle should be genuinely worried. i might be wrong about the speed here, but i don't think i'm wrong about the direction.
what to do about it
if you run a company, the question is not "how do we adopt ai." that question produces pilots, and pilots produce demos, and demos produce nothing. the question is coase's: which parts of this organisation exist only because coordination used to be expensive? every one of those parts is now negotiable, and your smallest competitor is already operating without them.
if you're a builder: the default size of a serious company just changed. you don't need permission to stay small. small is not the compromise anymore. small is the design.
legacy code doesn't get fixed the day it's diagnosed. it gets rewritten when someone ships the replacement. that's happening now, one three-person company at a time.