The flattening is not a forecast anymore. Gartner predicted that through 2026, 20% of organizations would use AI to flatten their organizational structure, eliminating more than half of current middle management positions, and the data suggests it is well underway. Manager headcount at public companies has fallen more than 6% since 2022, faster than the decline in executive roles or the overall workforce. Korn Ferry found 41% of employees saying their organization has cut management layers. And Gallup's span-of-control measure has climbed to 12.1 reports per manager in 2025, up from 10.9 the year before and nearly 50% higher than when it was first measured in 2013. MIT Sloan researchers found that in companies deploying AI at scale, spans have stretched toward fifteen.

The logic behind it is clean enough. A large share of a middle manager's week goes to reporting, coordinating, tracking, and status-passing. AI does that work well and cheaply. So the layer that existed to move information looks, from the top of the org chart, like a layer that no longer needs to exist. Cut it, widen the spans, bank the savings, move faster.

That reasoning is about half right, and the half it gets wrong is the expensive half. The middle layer performs two different functions that are almost impossible to distinguish from above, because they produce similar-looking artifacts. AI replaces one of them almost entirely. It does not touch the other at all. Companies that understand the difference flatten well. Companies that do not are cutting into something they cannot see and will not miss until it is gone.

Relay and translation are not the same job

The first function is relay. Information moves up, instructions move down, status gets aggregated, dashboards get assembled, updates get compiled and forwarded. This is genuine work and it consumes an enormous share of managerial time, and it is exactly the kind of work AI is good at. A system that can pull status directly from the work itself, summarize it, and route it to whoever needs it does not need a person in the middle retyping things into a slide. Gartner's own framing makes this explicit, noting that AI deployment allows for increased span of control by automating scheduling, reporting, and performance monitoring. If your managers are largely relay stations, the case for flattening is strong and you should take it. Nobody should defend the human relay.

The second function is translation, and it is a fundamentally different act. Translation is converting between things that do not have equivalents. A strategy stated at the top, improve margin without hurting service, has no direct executable form; someone has to turn it into specific choices for specific people in specific situations, choices the strategy never anticipated. That is not moving information. It is creating information that did not exist, by combining a general intent with local knowledge about what is actually possible here, with these people, this month.

It runs the other way too. A team's reality, this process breaks whenever the vendor is late, which is often, does not arrive at the top as a clean signal. Someone has to recognize that a scattered set of complaints and workarounds constitutes a pattern, decide it matters, judge how to frame it so it is heard, and carry the political cost of raising it. That is not reporting. It is interpretation plus judgment plus a willingness to be the one who said the uncomfortable thing.

From the top of the org chart, both functions produce similar-looking outputs: a manager, in meetings, generating updates and instructions. That surface similarity is the entire problem, because it makes the two look like one job that AI can now do. The relay part it can. The translation part is judgment under ambiguity, and no dashboard performs it.

What translation actually consists of

It is worth being concrete, because "translation" can sound like a soft word for something optional.

It is turning ambiguous intent into specific action. Every strategy is underspecified when it reaches the ground. Someone has to decide what it means for this team, this week, and take responsibility for that interpretation being wrong.

It is deciding what travels upward. Most of what happens in an operation is noise. Someone has to distinguish the signal, and this is a judgment call with real consequences: escalate too much and leadership drowns, escalate too little and problems arrive as crises. Nothing about AI makes this call, because it is not a question of information availability but of significance, and significance is contextual.

It is arbitrating conflicting priorities in the specific case. Leadership says quality and speed. The team can have one this week. Someone decides, and owns the decision. The higher you go, the less anyone knows the specifics required to make that call well; the lower you go, the less authority exists to make it. The middle is where those two curves cross, which is precisely why the layer exists.

And it is absorption. A functioning middle layer buffers, holding uncertainty so the team can work, taking the political friction so the work does not stop. Remove the buffer and the friction does not disappear, it lands directly on the people doing the work, which is why alignment tends to get worse rather than better after aggressive flattening. Korn Ferry found 43% of employees saying leadership alignment suffered after their organization cut layers, which is exactly what you would expect if the removed layer had been doing translation and everyone assumed it was doing relay.

The honest part: much of the flattening is correct

It would be wrong to read this as a defense of every management layer, and plenty of organizations genuinely are over-layered in ways that slow them down and serve nobody. Amazon's Andy Jassy framed his company's cuts around exactly this, saying rapid growth had added layers that slowed decision-making, and explicitly said the move was not primarily AI-driven. Layers accumulate during growth for reasons that stop applying, and removing them is legitimate organizational hygiene rather than a fad.

The relay work really was a large share of many management jobs, and AI really does perform it better. A manager whose week is genuinely dominated by compiling status and forwarding it is doing work that should be automated, and pretending otherwise protects a job rather than a function.

The trend is also less uniform than the headline numbers suggest, and it is worth being precise about that. Gallup's data shows that while the average span rose to 12.1, the median team is still five to six people, roughly two-thirds of managers oversee fewer than ten, and the jump in the average is driven largely by an increase in very large teams of twenty-five or more. So the picture is not every manager taking on more, it is a minority of extremely wide spans pulling the average up. The Bureau of Labor Statistics also still projects management occupations to grow faster than average this decade, which suggests the role is being reshaped more than erased.

So the argument is not that flattening is wrong. It is that flattening is a precision operation being performed with a blunt instrument, because the thing being cut contains two functions and the org chart shows only one.

What happens when you cut translation by accident

The failure does not announce itself, which is why it is dangerous. In the first months after a flattening, things often look better: costs are down, some decisions genuinely move faster, and the removed layer's absence is not obviously felt because the translation work does not appear on any report.

Then the second-order effects arrive. Strategy starts landing on teams in its raw, uninterpreted form, and people either freeze or invent their own interpretations, which diverge. Problems that used to get surfaced early arrive late and large, because the person who would have recognized the pattern and carried the political cost of raising it is gone. Conflicting priorities get resolved by whoever is loudest or most junior rather than by someone with the context and authority to arbitrate. And the remaining managers, now running twelve or fifteen people instead of seven, do triage instead of translation, because translation requires attention per person and attention per person is exactly what a widened span removes.

That last point deserves emphasis, because it is where the math turns against you. If AI absorbed only relay, then the remaining manager's job is now almost entirely translation, applied to twice as many people, with the same hours. Gallup's research is relevant here: manager engagement peaks at a relatively narrow span and declines as spans widen beyond it, with managers running very wide teams reporting higher workload demands and less control over their work. The work AI cannot do is the work that just got concentrated, onto people who are also absorbing new AI-related responsibilities without additional support. Flattening does not reduce the translation load. It redistributes it onto fewer people and then quietly hopes.

The test before you cut

The practical discipline is to stop asking whether a layer is expensive and start asking what that layer is actually doing, which comes down to a specific question you can put to any management position you are considering removing: if this person disappeared tomorrow, what decisions would go unmade, and who would make them instead?

If the honest answer is that some reports would stop being compiled and some meetings would stop being held, you have found relay, and you should cut it with confidence and let AI do the work. If the answer is a list of judgment calls, escalations, interpretations, and arbitrations that would either fall to someone without the context to make them or simply not happen, you have found translation, and removing it will cost you far more than the salary. Most positions contain some of both, which means the useful outcome of this exercise is usually not "keep or cut" but "cut the relay, and make sure the translation lands somewhere deliberate rather than nowhere."

This is the same discipline that applies to automating any role, distinguishing the visible tasks from the value that was never written down, explored in the job you're actually automating. Applied to the middle of an org chart, it has a specific shape: the layer was never really about moving information, and the companies that thought it was are about to discover what it was actually for.

FAQs

Q1. Is the flattening trend real or just headlines?
It is real and measurable. Gartner predicted that through 2026 one in five organizations would use AI to flatten structures, eliminating more than half of current middle-management roles, and manager headcount at public companies is down over 6% since 2022. Gallup's span-of-control measure has risen to 12.1 reports per manager in 2025, nearly 50% higher than in 2013, with spans stretching further in AI-heavy organizations.

Q2. What is the difference between relay and translation?
Relay is moving information: compiling status, forwarding updates, passing instructions down and reports up. Translation is converting between things that have no direct equivalent: turning ambiguous strategy into specific action, recognizing which ground-level signals matter and escalating them, and arbitrating conflicting priorities in specific situations. AI performs relay well and does not perform translation at all.

Q3. If AI handles coordination, why keep middle managers?
Because coordination in the administrative sense is only part of the job. The remaining part, interpreting intent, judging what matters, resolving conflicts in context, and absorbing friction so work continues, requires judgment under ambiguity rather than information processing. Removing the people doing that work does not remove the need for it, it just leaves it undone or pushes it onto people without the context or authority to do it.

Q4. How do I tell whether a specific role is relay or translation?
Ask what decisions would go unmade if that person disappeared tomorrow, and who would make them instead. If the answer is that some reports and meetings would stop, the role is largely relay and can be automated. If the answer is a list of judgment calls that would fall to someone without the context or simply not happen, the role is doing translation.

Q5. Isn't some flattening genuinely necessary?
Yes. Layers accumulate during growth for reasons that later stop applying, and many organizations are legitimately over-layered in ways that slow decisions and serve no one. The relay portion of management work really was substantial and really is better automated. The argument is not against flattening, it is against flattening without distinguishing which of the two functions you are removing.

Q6. Why does alignment often get worse after cutting layers?
Because alignment was substantially produced by the translation work that got cut. When strategy arrives at teams uninterpreted, people either stall or improvise their own readings, which diverge. Korn Ferry found 43% of employees reporting that leadership alignment suffered after their organization eliminated management levels, which is the predictable result of removing a translation layer while assuming it was a relay layer.

Q7. What happens to the managers who remain?
They get more of the work AI cannot do, spread across more people. If AI absorbed the relay portion, the remaining role is disproportionately translation, now applied to twelve or fifteen reports instead of seven, often alongside new AI-related responsibilities and without added support. Gallup's data shows manager engagement declining as spans widen, with very wide spans producing higher workload demands and less perceived control, which is why remaining managers frequently report strain rather than relief.

Q8. What should a leader do before approving a flattening?
Separate the two functions role by role rather than treating the layer as one thing. Cut the relay work confidently and let AI do it. Then decide explicitly where each piece of translation work will land, whether with a remaining manager who has capacity for it, with the team through changed decision rights, or upward. The failure mode is not cutting, it is cutting without ever deciding who now does the part that was never on the org chart.