You're Cutting the Wrong Layer

When the ATM arrived in the 1970s, everyone agreed it was the end of the bank teller. The machine did the teller’s entire job. Faster, cheaper, and it worked all night.

Then the number of tellers grew.

From roughly 300,000 in 1970 to almost 600,000 by 2010, even as banks installed some 400,000 machines.

The economist James Bessen worked out why. ATMs cut the number of tellers a branch needed from about 21 to 13. Cheaper branches meant banks could open far more of them. Fewer tellers per branch, many more branches, and the teller headcount went up for four decades.

We are terrible at predicting which part of a job a machine takes. We assume it takes the whole thing. It almost never does. It takes one layer and leaves the rest, and the part it leaves is usually the part that mattered.

I know this because I made exactly that mistake, in print, about a job I now think is the most misread role in the modern company.

A decade ago I argued that machines would replace middle management. The piece ran in Forbes. My case was clean: a software layer was learning to do what managers did, and it would do it without the manager.

The examples back then were the emerging gig platforms. Uber, Lyft, the design and cleaning marketplaces.

Each had built a piece of software that dispatched a worker, set the route, monitored the pace, handled the pay, and could fire you when your rating slipped. Management by API. Those were the early signals I saw at the time. The systems “managing” large workforces were nowhere near the capabilities we have today with frontier LLMs, but neither were the tasks and risk they were built to manage. I described the workers as cogs behind that API, and I asked whether the rest of us in management chairs were next (as systems got more sophisticated).

The coordination layer I described did get eaten. That part I got right. What I missed is that I’d confused two things that were never the same thing, and the confusion is about to cost a lot of companies a great deal of money.

Start with the part I got right, because the data is not subtle.

Middle managers were roughly a third of all layoffs in 2023, up from about a fifth in 2018.

That’s from Live Data Technologies, which tracks employment records for millions of workers. At small and mid-sized companies, the number of people reporting to each manager doubled between 2019 and 2024, from three to six, according to payroll data from Gusto. Across the whole economy, the average manager’s team grew to roughly 12 direct reports last year, up about 50% since 2013.

The names are familiar.

Amazon told its leaders to raise the ratio of workers to managers by 15% and hit the target early, mostly by merging teams and demoting managers into individual roles.

Google cut managers, directors, and VPs by about 10% in its efficiency push, and separately cut the number of managers overseeing fewer than three people by roughly a third in a single year.

Meta ran a “year of efficiency” that stripped out layers and pushed some engineering managers past 20 reports.

It even has a name now. The Great Flattening. And the people arriving in the workforce have already read the room: in one 2024 survey, more than half of Gen Z professionals said they don’t want to be middle managers at all, and most saw the role as high stress for low reward.

So yes. The layer is going. Score one for the version of me from ten years ago.

Take it as the whole story, though, and you’ll walk straight into an expensive mistake.

Peter Drucker saw this coming in 1988, and he described it more precisely than I did.

Writing in Harvard Business Review, he predicted that the information-based company would carry fewer than half the management levels of its day. Not because managing stopped mattering, but because of what most of those levels were actually doing. His words:

“Whole layers of management neither make decisions nor lead. Instead, their main, if not their only, function is to serve as ‘relays,’ human boosters for the faint, unfocused signals that pass for communication in the traditional pre-information organization.”

Human boosters for a faint signal. Sit with that, because it’s the whole thing.

Middle management was invented to solve a bandwidth problem.

When a railroad superintendent named Daniel McCallum drew one of the first modern org charts in 1855, he wasn’t in love with hierarchy. Trains were colliding on a single long line because nobody could move information up and down it fast enough. So he built layers. People stationed between other people whose job was to carry the signal.

The math forced it. A manager can only track so many interlocking relationships before they lose the thread, which is why span-of-control research from the 1930s landed on a handful of direct reports as the practical ceiling. Past that ceiling, you don’t get a better manager.

You get another layer. Every level of middle management since McCallum has been a workaround for the same limit: humans are a low-bandwidth network, so you insert more humans to move the messages.

AI is the first technology that actually fixes that. The signal no longer needs a human booster. Which means the relay dies, exactly as Drucker said it would.

But the relay was the cost of managing, not the point of it. The coordination was the tax you paid. The judgment was the job. Deciding what’s worth doing. Framing it so someone else can run with it. Knowing good work from plausible work. Building the capability of the person next to you. None of that was the relay. All of it survives.

The coordination was the cost. The judgment was the job. We spent a century confusing the two, and we’re about to find out how expensive that confusion is.

If you want to see where every management job is heading, watch the one that already arrived there. Software engineering.

Two years ago, an engineer’s core act was writing code. Today, at the companies moving fastest, it increasingly isn’t. More than a quarter of new code at Google is generated by AI and then reviewed and accepted by an engineer, per Sundar Pichai. Microsoft’s CEO has put the AI-written share of their code at 20 to 30%. The maker of the coding agent Devin says its own agents now open about a quarter of the company’s internal pull requests.

Look at what the human does in that sentence. Reviewed and accepted. Not authored.

The best engineers didn’t lose their jobs.

Their jobs turned into management. They decompose a problem, hand pieces to agents, set the intent, and then spend their time reviewing and verifying what comes back. GitHub’s CEO describes the advanced developer as someone who now “focuses on the delegation and the verification of a task,” working through “your orchestra of agents.” Anthropic runs a pattern internally where one lead agent breaks down a job and delegates to specialist sub-agents working in parallel. It is, quite literally, an org chart made of software.

McKinsey sketches the end state as a team of two to five people supervising an “agent factory” of 50 to 100 specialized agents. The engineer became the manager. The manager’s skills became the bottleneck.

And the tell that this is real, not hype: it doesn’t automatically work. A rigorous study by METR this year took experienced open-source developers, gave half of them modern AI tools, and measured the result. The AI-assisted group was 19% slower. Slower. Even more telling, those developers believed they’d been about 20% faster. The tool generated plenty. Judging and fixing what it generated ate the gains.

That’s not an argument against the shift. It’s the argument for it. When generation is free, the entire game becomes the quality of your delegation and your judgment. Bad managing of agents makes you slower. Good managing of agents is the skill. Same as it ever was with people.

So if you run a function and you can see this coming, what do you actually do? Not “embrace AI.” Something more specific. You start running your team, human and agent, the way the best engineers now run theirs. Five moves.

Decompose and brief. The scarce skill is no longer doing the work. It’s knowing what to ask for and describing it well enough that someone, or something, can execute. Ethan Mollick of Wharton points out that the artifacts good managers already produce, the product requirement doc, the creative brief, the clear set of instructions, are exactly what a capable agent needs. Your delegation documents are your prompts. If you can’t write a brief a smart new hire could run with, you can’t direct an agent either. This was always a management skill. Now it’s the management skill.

Design the evaluation before the work. Define what “done” and “good” look like up front, because you’re about to receive a firehose of plausible output and you need a way to judge it. When Microsoft asked workers which human skills matter most as AI takes on more, the top two answers were quality control of AI output and critical thinking. Not prompting. Judgment. The manager who can specify the acceptance test is worth more than the one who can only recognize good work when they trip over it.

Treat verification as the real work, not the afterthought. This is the hard one, and it’s where the engineers are bleeding. Generation collapsed to near zero cost. Checking did not. In one benchmark, an AI produced expert-level work in minutes that still took a human expert about an hour to verify. Your calendar has to invert. The time you used to spend producing, you now spend reviewing, and if you don’t budget for it, you’ll ship confident nonsense at scale.

Calibrate trust deliberately. Only about 6% of companies say they fully trust AI agents to run core processes on their own; most keep them on a short leash. That’s not timidity, it’s the job. Knowing which tasks an agent can own outright, which need a human check, and which shouldn’t be delegated at all is a judgment call you make dozens of times a day. It’s the same call a good manager makes about a junior employee, scaled up and sped up. You’re not managing whether the work gets done. You’re managing how much you trust it.

Orchestrate the blend. Your team is about to become people and agents in one unit, and your span of control is going to widen because of it. Microsoft calls the emerging role the “agent boss,” someone who builds, delegates to, and manages agents alongside their people.

Harvard Business Review argues we’ll see a formal “agent manager” role, the way “product manager” appeared during the software era. Whatever you call it, the shape is clear. Fewer people carrying signal. More people, at every level, orchestrating a mix of humans and machines toward an outcome.

Notice what’s on that list and what isn’t. There’s no new technical discipline you have to master. Every one of those five is a management skill you’ve been building for years. That’s the good news I got wrong a decade ago. The people best positioned for this aren’t the prompt engineers. They’re the operators who already know how to get work done through other people. The talent was never in the relay. It was in the judgment, and judgment is exactly what’s now scarce.

There’s one line in my old piece I’d write again today, though not for the reason I first wrote it.

Back then, worrying about the gig workers behind the API, I said they were entering a dead end, with little chance to progress, learn, or build skills that would pay off later. I meant it as a concern about Uber drivers. It turned out to be a concern about knowledge work, and it’s the crack in this whole optimistic story.

Because everything I just described, the manager as orchestrator of humans and agents, assumes you have managers. It assumes you have people with enough judgment to know when the agent is wrong. And judgment doesn’t arrive by download. It’s built on the bottom rung, doing the grunt work, being junior in a room full of seniors, making the small mistakes that teach you what good looks like.

That rung is exactly what AI is eating first. LinkedIn’s chief economic opportunity officer put it plainly: “breaking first is the bottom rung of the career ladder.” Stanford researchers, using payroll data, found employment for workers aged 22 to 25 in the most AI-exposed jobs has fallen 13 to 16% since generative AI took hold, while their older colleagues in the same roles held steady or grew. The junior tasks, the first draft, the initial analysis, the starter code, are the ones the agent does now.

Mollick names the paradox cleanly: juniors use AI the most but learn from it the least, while seniors extract the real value because they have the experience to catch what it gets wrong. You can’t become senior without being junior. And we’re quietly deleting junior.

I want to be honest about the uncertainty here, because the doom version of this is overcooked. The cause is genuinely contested. Yale’s Budget Lab, looking at the data 33 months in, concluded AI is “probably not yet” the reason the entry-level market is soft. Other economists pin most of it on remote work and the hangover from pandemic over-hiring, and point out that entry-level hiring started sliding before ChatGPT existed. The collapse of the training ladder is a risk we can see the mechanism for, not a catastrophe we’ve measured. Yet.

But the mechanism is real, and a few companies are already treating it as one to manage rather than ignore.

Some are pushing judgment-heavy work down the ladder sooner, handing new hires the complex, interpretive tasks that used to require years of tenure, precisely because the AI now handles the rote parts they used to cut their teeth on. KPMG has new grads doing higher-level tax work. Firms are experimenting with deliberate senior-junior pairing, using the AI to capture the reasoning so the work becomes teachable. The move is to stop treating entry-level hiring as a cost to be optimized away, and start treating it as what it actually is: an investment in the future stock of judgment inside your company.

So I called the funeral ten years too early, and I stood at the wrong grave.

The relay is dying. Good. It was always the tax, never the point, and Drucker was calling it a redundant human booster before most of today’s managers were born. What’s left when you clear it away is the real work, deciding what matters, briefing it, judging it, building the person who’ll do it next. That work isn’t shrinking. It’s becoming everyone’s job.

The danger was never that managers become obsolete. It’s that we get so efficient at removing the layer that we take out the ladder with it, and wake up in a decade with an org chart full of orchestrators and no one who ever learned to play.

A manager’s real product was never the status update. It was the next manager. Delete that by accident, and the whole thing will look wonderfully efficient right up until the moment it stops working.