Most of the conversation about AI and jobs centers on a natural question: which tasks can AI perform? That question matters, and a substantial body of work addresses it. But task replacement may be an incomplete lens for understanding how organizations actually change.
The task-replacement lens holds the organization fixed. It takes a company (same structure, same reporting lines, same division of labor) and asks which boxes on the org chart get handed to a machine. Swap out the human, slot in the algorithm, everything else stays put.
But recent evidence suggests firms are also reorganizing.
Ewens and Giroud (2025) studied over 3,100 U.S. public firms and found that after adopting AI, these companies flattened their management hierarchies. Not a little. Meaningfully. Layers of middle management compressed. Spans of control widened. Babina et al. (2025) found the same thing. These firms did more than automate a few tasks. They changed how work gets organized.
The mechanism behind this may be coordination compression: AI reducing the overhead of organizing people. In a recent paper, I formalize this channel and trace how outcomes depend on two organizational parameters.
Much of the AI-and-work literature focuses on AI performing tasks currently done by humans. Writing code, drafting memos, analyzing data, generating images. This task-substitution channel appears in headlines, policy debates, and most economic models.
That channel is real and well-studied. This paper examines a second channel that may also reshape organizations.
Call it the coordination channel. Every organization pays a cost, beyond salaries, just to keep people aligned. Communication between teams, monitoring progress, scheduling work, turning strategy into plans, routing information up and down the chain. Managing is coordinating. And that coordination has a cost per person managed.
You might be thinking: hasn't every IT upgrade since the telephone reduced coordination costs? Email made communication faster. Slack made it real-time. Shared documents killed version confusion. All true. But there's a difference between making information move faster between people and removing the cognitive work those people exist to do.
Email lets a manager hear from ten reports faster. The manager still has to read all ten messages, figure out what matters, decide what to do, and respond. Slack tightens that loop, but the loop is still there. The manager is still the bottleneck, still doing the processing. Those tools made the pipe wider. They didn't change how many people you needed to process what came through it. These tools did not produce the kind of hierarchy flattening observed in recent studies of AI adoption.
AI can do some of the processing. It reads the ten updates, flags the two that are behind schedule, drafts responses for the eight that are on track, and proposes a reallocation for the two that aren't. The manager reviews the output instead of building it from scratch. In principle, that changes how many people you need in coordinating roles.
When AI drives coordination costs down this way, the math changes. A manager who could previously oversee eight people can now oversee fifteen or twenty. The expansion comes from AI absorbing coordination load rather than from any change in the manager's underlying ability. It summarizes updates, flags problems, schedules across time zones, keeps track of what twenty people are doing at once. The manager's reach extends. Their span of control widens.
The task channel and the coordination channel address different questions. One asks which jobs are affected by automation. The other asks which organizational layers are affected by cheaper coordination. The answers, and the people involved, may differ substantially.
Think about what a middle manager actually does all day. A huge chunk of it is coordination. Passing information up and down the chain. Checking on project status. Turning executive priorities into concrete assignments. Pulling together reports. Sorting out dependencies between teams. When AI picks up that coordination work, the managers above don't need as many people between them and the front line. One senior leader can now cover what two or three middle managers used to handle.
The formal version: each manager's span of control is the inverse of their coordination cost. As agent capital rises, coordination costs fall, and spans expand. The formula is direct.
Si = (1 + ฮณ · KA · siฮฒ) / c0
In this framework, the middle manager's role compresses because AI absorbs the coordination that justified the position. The layer thins because the work above and below it can be organized with less intermediate oversight.
This is consistent with what early evidence shows. Firms appear to be flattening because AI made the coordination those managers performed something that can be folded into other roles.
So coordination gets cheaper. Hierarchies flatten. Is that good or bad?
It depends. And the answer turns less on the technology itself than on two organizational parameters.
Who captures the coordination gains?
When AI makes coordination cheaper, managers can oversee more people. But the question is whether every manager benefits about equally, or whether the gains pile up at the top.
If every manager gets similar leverage from AI coordination tools, the benefits spread out. Spans widen across the board. The hierarchy flattens more or less evenly. Everyone's a bit more productive.
But what if AI gives your best managers a much bigger boost than everyone else? What if your top performers can now manage fifty direct reports while your average managers go from eight to twelve? The absolute numbers aren't that different. The gap between them is enormous. Same tool. Same rollout. Wildly different results.
This is where ฮฒ enters. It's a single parameter that captures how concentrated the coordination gains are. When ฮฒ is low (think electricity), everyone benefits roughly equally. When ฮฒ is high (think a Stradivarius violin), only virtuosos extract extraordinary value from the instrument.
The skill gap extends beyond technical fluency. Some people use AI to rethink how their entire team works. They redesign workflows, eliminate unnecessary roles, collapse process that existed only to move information around. Others use it to write emails faster. The difference in leverage between those two uses isn't small. It's easily ten to one. The difference is skill, not access. Giving everyone the same tools doesn't mean everyone gets the same mileage out of them.
Does cheaper coordination create new work?
When coordination gets cheaper, some things that were previously too expensive to organize become worth doing. A team of three that would have needed its own manager, too small to justify the overhead, can now be tucked under an existing manager's wider span. New projects, new functions, new roles that never made economic sense before can now exist.
The parameter ฮด captures this. When ฮด > 0, the task frontier expands as coordination costs fall. When ฮด = 0, the organization just does the same things with fewer people.
These two forces, who gets the gains and whether new work shows up, produce four different worlds.
Click any quadrant to explore what happens under each combination of ฮฒ (who benefits) and ฮด (whether new work emerges).
Adjust ฮฒ and ฮด to see how coordination compression plays out across managers with different skill levels. The chart shows how each manager's span of control changes.
The coordination compression framework rests on assumptions worth scrutinizing. Here are the strongest objections, taken seriously.
Most of AI's impact could still be plain old task substitution. If what AI mostly does is write emails, generate code, and summarize documents, all task-level stuff, then coordination compression is a sideshow. An interesting sideshow, but a sideshow. Most economic models focus on the task channel for a reason. That's where the measurable impact is clearest.
But the evidence says the coordination channel is already big enough to see. Ewens and Giroud didn't study a handful of firms. They studied 3,100 public companies and found systematic hierarchy flattening after AI adoption. That's not email drafting. Something structural changed in how these firms organize people. Task substitution alone doesn't explain why the shape of the organization changes. It only explains why certain positions disappear. If AI just substituted for tasks, you'd expect the org chart to keep its shape with a few boxes gone. Instead, the chart itself is changing. That's the coordination channel at work.
Hierarchies persist for political reasons, not efficiency reasons. The VP of a division doesn't lose their reports because a paper says coordination costs fell. There's enormous resistance to restructuring. Managers who'd lose their positions. HR systems built around fixed levels. Cultures where seniority means headcount. The theory might be right in the long run, but the friction could take decades to work through.
I don't dismiss inertia. It's real. But the firms in these studies didn't flatten because they read my paper. They flattened because competition made the old structure too expensive once the technology made a new structure possible. Inertia slows things down. It doesn't stop them. And the pressure is only going to increase as early movers show what flatter, AI-augmented organizations can do.
AI tools are cheap and getting cheaper. Eventually everyone has access. The concentrated-gains scenarios become irrelevant. Give it a few years. The tools commoditize. The skill gap closes. The only quadrants that matter long-term are the ones where gains are broad.
This objection deserves careful examination. Access is not the same as leverage. Universal access to a piano doesn't make everyone a pianist. The gap between "I use AI to draft emails" and "I use AI to rethink how my whole team works" is a skill gap. It comes down to how well someone understands their own organization's workflows and where the coordination bottlenecks actually are. Making the tools free doesn't close that gap. If anything, universal access can widen it, because the people who already know how to use these tools well benefit the most from having even better ones.
Where AI takes your organization depends on the choices around it. Who gets the tools. How they're trained. Whether the gains go toward growing what the organization does or just cutting what it spends.
Not entirely up to you, of course. Market forces and the pace of change constrain the options. You can't will yourself into Rising Tide if your industry is contracting. But within those constraints, there's real room to move. Training people broadly rather than concentrating AI tools among a few top performers pushes toward shared gains. Building new roles rather than just cutting old ones pushes toward a growing pie rather than a shrinking one.
This framework doesn't only apply to traditional organizations. If you're a solo operator running multiple AI agents, one writing code, one handling design, one doing research, one managing deployment, you're a manager. Your agents are your reports.
You face the same coordination costs. You have to give each agent the right context, make sure one agent knows what another already did, verify outputs, decide when to trust them and when to check. Every improvement that makes that easier (better context windows, persistent memory, shared tool access) is coordination compression. The math is the same.
And the regime fork applies here too. Someone who deeply understands software architecture gets far more out of the same set of coding agents than someone who doesn't know what to ask for. The agents amplify judgment, not just effort.
The paper develops all of this formally. Five propositions, each proven mathematically. Here's what they say in plain language.
The math formalizes an intuition that can be stated simply: whether AI benefits workers broadly or concentrates gains depends less on the technology than on which regime applies. The relevant question is whether the coordination savings spread or concentrate, and whether new work emerges. Unlike the technology itself, those factors are at least partly within organizational control.
The full treatment, with proofs, simulations, and robustness checks, is at arxiv.org/abs/2602.16078.