I keep seeing the same line in threads about AI and org charts. Someone says it cleanly, almost with relief: middle management is just a dumb relay layer. People who take a number out of one box and type it into another. Useless. AI will wash them out.

It’s a satisfying thing to say. It’s also half wrong - and the half that’s wrong matters more than the half that’s right.

Management layers did not appear by accident. The Roman army had a hierarchy. Not a deep one, but a real one. And it wasn’t there because some Roman consultant sold them an org chart. It was there because it solved a problem that has not gone away.

The layer the Romans couldn’t skip

Pull apart a Roman legion and you find a structure that looks almost modern.

Eight men formed a contubernium - the squad that shared a tent and a cooking pot, with an experienced soldier, the decanus, in charge when no orders came from above. Ten of those made a century: around eighty fighting men under a centurion. Six centuries made a cohort, roughly 480 men. Ten cohorts made a legion, about 5,000.

Count the layers. Legate, then cohort, then century, then contubernium, then the individual soldier. Four, maybe five levels of command for five thousand people under arms. A modern company of five thousand routinely runs eight to twelve.

The Romans weren’t being clever here. They were being forced. A commander can hold only so many things in his head at once. He can give orders to a handful of subordinates, watch them, correct them, and still have attention left for the enemy. Past that handful, he can’t. So you insert a layer. Not because the layer is smart, but because the alternative - one man directing five thousand - is physically impossible.

We have a name for this now: span of control. And there’s an old, slightly grim piece of math behind it. In the 1930s a management theorist named Graicunas worked out that when you add subordinates to a manager arithmetically, the number of relationships that manager has to track grows geometrically. It isn’t just you-to-each-person. It’s the relationships between them, and between every subgroup of them. Add the sixth direct report and you haven’t added one relationship. You’ve added dozens.

The centurion existed to absorb that geometry. And here is the part the “dumb relay” framing misses completely: the centurion was not a relay. He had the hardest job in the legion. He drilled the men, enforced discipline, and stood in the front rank when the line met the enemy. Roman writers called the centurions the backbone of the army - career soldiers, the ones who actually held a battle together when it started to go wrong.

That layer, the one that exists because coordination has a hard ceiling, is not washing out. AI does not repeal span of control.

You can watch modern organizations rediscover the same ceiling. Amazon, where I spent years working on AWS SDKs and reliability, built a whole doctrine around it - the two-pizza team, small enough to feed with two pizzas, with a single owner accountable for it end to end. And one level up, a Senior Manager would carry 70 to 100 people - through intermediate leads, but with full accountability for every one of them. Set that number next to the centurion’s eighty. Two thousand years apart, the same human ceiling, and the same demand under it: one named person who answers for the whole group. That’s the Roman insight with a catering metaphor - past a certain size, a group needs an internal seam, and that seam needs an owner. Nobody at Amazon would call a two-pizza lead, or that Senior Manager, a relay.

The layer the Admiralty couldn’t stop

Now the other story.

In 1955 a British naval historian, C. Northcote Parkinson, published a short, funny essay that became famous as Parkinson’s Law: work expands to fill the time available for it. The example he built it on was the British Admiralty.

His numbers: in 1914 the Royal Navy had 62 capital ships and about 146,000 sailors, administered by 2,000 Admiralty officials. By 1928 the fleet had shrunk to 20 capital ships and 100,000 sailors - and the number of officials had grown to 3,569. Fewer ships, fewer sailors, far more administrators. Parkinson’s point: the office grew on its own, decoupled from the work. Officials hire subordinates, not rivals. Officials make work for each other.

That is the parasitic layer - the one that grows because organizations grow, regardless of whether there is anything new to coordinate. This is the layer the “it’ll wash out” crowd is actually describing. And they’re right about it. It should wash out.

But I want to be honest about Parkinson, because the real story is more interesting than the slogan.

Parkinson’s essay was satire, and his statistics don’t fully survive scrutiny. Later naval historians pointed out that his 1914 figure was too high - which, awkwardly for the critics, makes the real growth larger, not smaller, closer to 120% than 78%. The deeper problem is what Parkinson left out: there was a world war in the middle of his timeline. Between 1914 and 1928 navies acquired aviation, submarines, radio, signals intelligence. Some of that Admiralty growth wasn’t entropy at all. It was real new coordination work for a fleet that had genuinely become more complex, even as it got smaller.

That isn’t a flaw in the argument. That is the argument.

The hard part was never “do layers grow.” The hard part is telling the two kinds of growth apart. Some of the Admiralty’s new officials were pure Parkinson - boxes that existed to justify other boxes. And some were the modern equivalent of a centurion for a domain that genuinely hadn’t existed in 1914. From outside, sitting in the same building, wearing the same title, they looked identical.

What AI actually eats

This is where the AI conversation usually goes wrong.

The popular version: AI does execution, so the execution layer dies, and management - the coordination layer - is safe. That’s comfortable, and it’s already out of date. Agentic systems don’t just execute. The whole orchestration pattern - a top-level agent routing work to sub-agents, passing context between them, handling retries and handoffs - is coordination. AI is moving into the coordination layer too.

So the line is not execution versus coordination. If you defend “coordination” as a category, you’re defending something AI is actively learning to do.

The real line runs somewhere else. It runs between mechanical transmission and judgment.

Mechanical transmission is taking a status from one place and rendering it in another. Aggregating five updates into one. Routing a task to whoever is free. Translating a request into the format the next system expects. This is the relay work, and structurally it’s the same thing as a human sitting between you and a piece of tax software, typing your numbers into the right boxes. AI is very good at this. It will take this work, in management and everywhere else. The dismissive take had this part right.

What AI does not take is the centurion’s actual job. Judgment under uncertainty, when the data is incomplete and contradictory. Accountability - being the named person who answers for the outcome, not for the routing of it. The escalation call: this one is fine, that one has to stop now. Reading people - knowing that this engineer is two weeks from burning out and that one is quietly bored. Holding the line when a plan meets reality and starts to come apart.

None of that is transmission. All of it is what span of control actually exists to provide. The centurion’s value was never that he carried messages between the legate and the soldiers. It was that he could be trusted to hold eighty men together when the plan failed. That is still a job. It may be the only management job that’s left.

The megamanager bill

There’s already evidence of what happens when you cut without making this distinction.

Companies are flattening hard. Gartner expects a large share of organizations to use AI to remove management layers, and postings for middle-management roles have dropped sharply since 2022. But look at what happened to the managers who stayed. The average American manager now has around twelve direct reports - roughly double what it was a decade ago. And in one survey, a third of HR leaders said AI-driven restructuring had stripped out institutional knowledge the remaining staff simply could not replace.

That’s the bill for cutting with an axe. When you flatten without asking which layer you’re removing, you don’t remove the Admiralty layer cleanly. You remove some of it, and some of the Roman layer with it, and you hand the survivors a span of control the Romans would have called impossible. The geometry doesn’t care that you have AI now. Twelve people is twelve people’s worth of relationships, and a manager drowning in them isn’t doing the centurion job either. They’ve just become a more expensive relay.

Which layer are you

So the slogan needs its second half. The dumb relay layer will wash out - yes. But middle management was never only the relay layer. It was always two things in one suit: the centurion and the Admiralty clerk. One is a load-bearing wall. The other is a wall built to keep the first wall company.

The useful question isn’t whether layers survive. It’s the one you can ask about your own job, or the one above you, tomorrow morning. Strip out everything that’s transmission - the status reports, the aggregation, the routing. When that’s gone, is there a centurion underneath?

If there is, AI just took your busywork and left you the hard part.

If there isn’t - if transmission was the whole job - then the relay was the job, and you already know how that story ends.