TL;DR: Organizations are shaped the way they are because no single human head can hold or process everything a company knows. Hierarchy, middle management, layers: all of it exists to work around that limit. AI removes the limit, which makes the entire shape optional for the first time. But you can't get to the new shape by adding AI to the old one; the old structure has to come down before something different can be built in its place. So real transformation runs: shrink, rebuild around AI, then expand into ambitions the old structure could never attempt. Almost no one is doing this. The absence of real shrinkage today means almost no one has started. The alternative isn't a gentler version of the same outcome; it's what happened with electrification: thirty years of marginal gains before anyone captured the real prize.
Contents: Humans as Routers · Why Organizations Look the Way They Do · AI Attacks the Cognitive Strand · The Cycle: Shrink, Rebuild, Expand · Incumbents or New Logos?
There is a popular story about AI and jobs: the machines are getting smarter, they will do the work, and so people get cut. Here's the inconvenient fact underneath it: outside of tech, the layoffs mostly aren't happening 1. Most people read that as reassurance: maybe the disruption was overhyped. There is a possible opposite read. Real AI transformation requires workforce reduction as its first act, not its side effect. The layoff is not the automation, it is the ground-clearing before a rebuild, and the absence of that reduction tells you that almost nobody has actually started. What follows is the argument for why the rebuild must run this way, on a five plus year horizon, and why the alternative is decades of marginal gains.
The middle of most organizations takes information from one place, filters it, summarizes it, translates it, and passes it somewhere else. Status reports go up. Priorities come down. Requirements move sideways. Most middle roles, and many senior ones, are fundamentally routing roles: humans moving knowledge between other humans. The same is true of the software stack underneath them: most enterprise software is clerical routing encoded as forms, approvals, tickets, and handoffs.
Jack Dorsey described his new organizational structure as humans on the edges 2, not in the middle. The edges are where judgment, relationships, taste, and accountability live. The middle is where information gets moved around. AI takes the middle. What's left for people is the edges, and the edges are the good part.
But why is the middle shaped this way, in every company, every industry, every country? The answer explains why AI hits the middle first and hardest.
Organizations exist in their current shape for a braid of three reasons: economic, cognitive, and social. AI does not attack them equally.
The economic strand is the oldest argument. Why do companies exist at all if markets are efficient? The answer was that contracting every task on the open market is expensive (finding people, negotiating, monitoring), so it's often cheaper to hire people and coordinate work inside a firm. Firms exist where internal coordination beats market transactions. True, but it only explains why the firm has a boundary. It doesn't explain the shape inside the boundary: why every large organization, in every industry and country, converges on the same pyramid. 3
The cognitive strand explains the shape. A human being has finite attention, memory, and processing capacity, so no one person can hold everything an organization knows or track everything it does 4. Hierarchy is the machine we built to work around the limits of one head. A manager can effectively direct only about five to eight people before monitoring and context-holding saturate their bandwidth, and that single number generates the entire pyramid by arithmetic: ten thousand workers at a span of seven forces four to five layers of management into existence whether anyone wants them or not. And what do those layers do all day? Economists like Luis Garicano showed that they are knowledge hierarchies: routine problems get solved at the bottom, exceptions escalate upward to people who hold rarer knowledge, and each layer economizes on how much any one person must know. The middle of the organization is, quite literally, an information-processing structure made of people, which is why it behaves like the routing layer described above. The org chart was never a design choice. It was a workaround for the human head. 5
The social strand then grows on top: careers, status, and pay all get anchored to the structure the cognitive limits created. Promotion in a coordination-built firm rewards coordinating well (not breaking things, managing upward, keeping the machine smooth), and the people who rise on those criteria set the criteria for the next generation, compounding the filter for decades. Compensation locks it in: organizations pay for scope, and scope is measured in headcount, so a VP's importance is the size of their pyramid.
Three strands, one shape. But they are not equally vulnerable.
AI attacks the cognitive strand first and most directly, because before it is anything else, AI is an information-processing technology, and before it is anything else, an organization is an information-processing structure. The constants that generated the pyramid stop being constants. An LLM's attention doesn't saturate at seven direct reports. It holds the exception-handling knowledge that middle layers existed to store. It aggregates and routes natively, in parallel, across the whole organization at once. When the cognitive constants change, the arithmetic changes, and the pyramid is no longer implied. The other two strands unravel as consequences: coordination costs collapse because cognition got cheap. One high-agency person directing AI systems does what previously took a team plus a manager plus the reporting structure around them, and the status system loses its measuring stick, because headcount stops implying capacity.
Two things follow from the collapse. The firm can produce today's outputs with far fewer people. That's the shrink. And the old ceiling on ambition moves. Complexity was always capped by what could pass through chains of bounded human processors, so when the bound lifts, organizations can attempt things that were impossible at any headcount. That's the expansion at the end of the cycle.
But the social strand doesn't just unravel. It resists, and this is why the shrink can't be finessed. A firm that spent decades promoting for coordination discovers there are not enough high-agency people inside to execute a rebuild, and no quick fix exists: you can teach someone a tool in a month, but you cannot quickly un-teach two decades of learned deference to process. Worse, in a system that no longer needs coordination, coordination roles turn negative value: their instinct is to reinsert the checkpoints and approvals that justified them, imposing process on people who no longer need it. And the people who must execute the shrink are precisely the people whose status it destroys. That, more than any technical limitation, is why transformation stalls.
One clarification. None of this has much to do with experience or age: deep domain judgment (knowing what good looks like) becomes more valuable in the new model, not less. What evaporates is procedural expertise: knowing whose approval to get, which forum to raise things in, how to navigate the org as it exists. Call it the difference between knowing the syntax of the organization and knowing its semantics: the forms versus the meaning. The syntax moat disappears overnight; the semantics compound.
Phase one: shrink. Not because AI took the jobs, but because you cannot rebuild an organization's information flows while thousands of people are actively maintaining the old ones. Every existing process has defenders, and the defenders are not malicious. They are doing their jobs. The ground has to be cleared.
If the shrink is a precondition rather than a consequence, then the current quiet in the labor market isn't evidence against the thesis. It's evidence that phase one has barely begun. And it explains why tech firms are cutting first: they are closest to the technology, they build the substrate everyone else will run on, and they face the most pressure to model the new shape. They aren't proving the whole economy is transforming. They're running phase one early. The shrink also won't always arrive as a dramatic layoff announcement. Attrition without backfill, hiring freezes, headcount held flat against growing revenue: these are quieter versions of the same ground-clearing. The leading indicator to watch is revenue per employee becoming the number boards brag about.
Phase two: rebuild. Re-implement what the organization does on an AI substrate: same outputs at first, radically different internals. Agents and tools in the middle, humans at the edges driving the loop.
Phase three: expand and rehire. With the new architecture in place, the organization's ambition can grow far beyond what the old one could attempt. Hiring resumes: some adapted incumbents, and a new cohort of people who never knew the old way of working, the same way developers who grew up on the cloud never learned to rack servers.
The obvious objection to phase one: doesn't ground-clearing destroy tacit knowledge? Yes. The routing humans were also doing invisible error-correction, and some of it was load-bearing. But AI changes the economics of this trade-off: reconstructing process knowledge in an AI-rebuilt system is now cheaper than fighting process defenders in the old one. It is a real cost, but a survivable one, and smaller than the cost of not rebuilding.
Why five plus and not two? Because we have run this experiment before. When factories electrified, the ones that bolted electric motors onto steam-era layouts got marginal gains for decades. The productivity explosion only came when a new generation rebuilt factories from scratch around distributed power, and that took roughly thirty years, and often new companies. The bet embedded in a 5+ year timeline is that software organizations reconfigure faster than physical plants. That seems right, but it is a bet.
The electrification story also tells you what the alternative path looks like, because it is the default. If organizations refuse the rebuild, if they keep bolting AI onto intact structures, the AI future still arrives, but as the slow path: thirty years of marginal gains, generational turnover doing the work that decisions could have done, and most of the opportunity simply forfeited to whoever eventually builds fresh. The shrink-rebuild-expand cycle isn't inevitable. It's the fast path. The choice organizations face is not whether to be disrupted, but whether the transformation takes one decade or three.
Will existing organizations actually complete the shrink-rebuild-expand cycle, or will they be out-competed mid-rebuild by new entrants who skip the rebuild entirely and start natively in the new model? Amazon did not wait for Sears to finish digitizing.
It depends on the specific case, and I'm deliberately not predicting winners. The cycle holds at the level of the economy either way. Either the incumbents transform, or they shrink and die while new organizations are born already rebuilt, in which case the "rehiring" phase is simply labor reallocating across firms rather than within them. The macro direction is the same. The logos are an implementation detail.
In next few posts will explore where we are today and how the substrate seems to be evolving.