About this series: For decades, companies have paid for two entirely different things under the same word: coordination and judgement. AI makes one of them almost free and the other one scarce. Part 1 asks what that means for people and careers. Part 2, what it means for structure and speed. Part 3, why missing judgement in leadership roles no longer merely slows things down but does damage.
Volkswagen, BMW, Bayer, Deutsche Bahn, BASF, SAP, Evonik. Within a few quarters, all of them announced the same thing: fewer management layers, fewer management positions. Executive search firms now report the same movement from the German mid-market.
The usual explanation is cost pressure. And that explanation is correct.
At Volkswagen it is about falling sales, China, the cost of electrification and a return that is meant to be back in double digits by 2030. AI appears nowhere in those justifications. Bayer halved its management layers from twelve to six because an operating model called for it, not because a model writes reports. The restructuring did not begin with AI, and it would be happening without it.
And yet something is different from every round of cuts before this one.
Until now it always went the same way: a corporation cuts the middle layer during a crisis, and three to five years later it is back. Under a different name, in a different colour on the org chart, but back. Because the work that sat there had not disappeared. Someone still had to report, align, translate and allocate.
This time, the way back is not part of the plan. And for the first time, part of that work actually has a substitute.
That is where the real problem starts. Because what sat in that layer was never one thing. It was two, and they were never separated, because they never had to be.
Coordination. And judgement.
One of them can be replaced. The other cannot. And anyone who cuts both together only notices the difference when it is too late.
What agents take over first
A large share of a manager's week consists of reporting, aligning, reviewing and translating between layers. In other words, moving information around. Practitioner accounts put that share at roughly 60 percent; no robust primary survey exists. What is not disputed is the direction: these are exactly the activities that automate first, and they do so cleanly, cheaply and around the clock.
As early as 2024, Gartner forecast that by 2026 one in five organisations would use AI to eliminate more than half of its existing middle management positions. Whether the ratio holds is open, and the forecast describes an intention rather than a measurement. But it describes precisely why nobody is planning for a rebuild this time.
That also breaks the span of control. Seven to ten direct reports was the long-standing rule of thumb. How far spans genuinely widen under broad agent use is empirically inconsistent; considerably higher figures circulate in the debate. The mechanism is clear regardless: anyone scaling with agents needs fewer people to pass information along.
Fewer layers. Wider spans.
And a leadership role whose content is something entirely different from before. Because what remains is not the easy part. It is the hard part: deciding exceptions, resolving conflict, developing people, steering mixed teams of humans and agents.
That is a rarer capability than condensing status reports. And one that almost no company has ever trained systematically.
The error in the argument about generalists
Few topics currently produce more contradictory writing. One camp says AI rewards the versatile generalist who now fills five roles at once with the new tools. The other says the market only pays for deep specialisation.
Both camps talk past each other, because they use the same word for two entirely different people.
There is the broker. His value comes from knowing who to ask, how to phrase it and where things usually get stuck. He knows the frame of every topic and the floor of none. Breadth without depth.
And there is the practitioner with reach. Someone who was genuinely deep in at least one field, long enough to have made mistakes there and lived with the consequences. From that base, they built breadth.
AI eats the first type. Not eventually. Now.
Why? Because the broker's core service is producing plausible connectivity on any topic. That is exactly the discipline in which a language model is unbeatable at marginal cost near zero. A model is the perfect generalist without depth. It can speak to everything and is responsible for nothing.
Anyone whose role rests on the same service is not competing with a tool. They are competing with a copy of themselves that is faster and costs nothing.
Judgement is the ability to tell correct from plausible
This is the actual point.
AI systems rarely produce obviously wrong answers. They produce convincing ones. The difference between a sound answer and a dangerous one is usually invisible in the output.
It is visible only to someone who has done the thing themselves.
Overview says: that sounds coherent.
Judgement says: that sounds coherent, and it is still wrong here.
That second capability does not come from reading widely. It comes from experience with consequences. Someone who has never carried a decision that went wrong has no material against which to calibrate. They can only assess results by checking whether they feel right. That is not assessment. That is taste.
This is exactly why the notion of the player-coach is returning: whoever leads has to be deep in the substance, in the same way that nobody runs a team at a law firm without mastering case law. The link between promotion and team size is also dissolving at some companies. The people who rise are not those who lead many, but those who achieve much with few.
Leadership is becoming substantive again. Not because the human side of leadership matters less.
But because the content-free version of leadership has become automatable.
And there is a hard flip side: anyone who does not master the new tools themselves becomes the bottleneck for their entire team. The contrast between two surveys is telling here. In a study by Stifterverband and McKinsey, 79 percent of around 1,000 German executives said their employees lacked basic AI skills. The finger points downwards. In the global Roland Berger survey, by contrast, 59 percent consider their own leadership level insufficiently prepared.
So the gap does not only sit at the bottom. That is just where it is usually looked for.
The problem almost nobody talks about
Depth has a supply chain. And it is being cut in two places at once.
First, at the bottom. The Stanford Digital Economy Lab, together with ADP, analyses monthly payroll records of millions of US employees. In the version updated in August 2026, employment among 22- to 25-year-olds in highly AI-exposed occupations sits 19 percent below the level it would have reached had it grown in line with less exposed peers. There is no comparable gap among experienced employees. Two caveats belong with this: the authors explicitly treat it as a descriptive early indicator, not proof of causation. And the effect runs through fewer hires, not through layoffs.
In Germany, the share of advertised entry-level positions in 2025 was around 42 percent below the average of the previous five years, based on an analysis of 4.6 million job ads at Stepstone. The IAB attributes that drop primarily to the economic cycle rather than to AI. The structural cause comes from the Randstad-ifo HR survey for the fourth quarter of 2025: at 14 percent of companies, AI already handles tasks that would normally have gone to entry-level staff. For the coming three years, 40 percent plan to do this, rising to 63 percent among large companies.
Second, in the middle. The middle layer was never merely administration. It was the stretch on which people learn to carry responsibility before the sums get large. Cut it in 2026 and you will search in vain for successors with leadership experience in 2028.
The result is a market looking for something whose production it discontinued itself: juniors with senior profiles.
They will not appear. You cannot buy judgement once nobody provides the years in which it forms.
What leadership has to make of this
Four consequences that, in my view, will decide the next few years.
Recut the work, do not just cut the positions. Cutting layers without first establishing which decisions were actually made there does not relocate those decisions. It lets them lapse. That only becomes visible two quarters later, in the form of mistakes nobody caught in time.
Rebuild entry roles around assessment. The old junior role consisted of support work. The new one has to consist of review: assessing AI output, finding errors, being able to explain why something does not hold. More demanding than before. But it is the only version of an entry role that produces a populated senior bench in five years.
Do the honest maths on spans of control. Gallup reports a marked drop in manager engagement for 2025, from 31 percent in 2022 to 22 percent, while the teams they lead grow larger. Notably, the strongest driver of genuine AI impact within a team is active support from the direct manager. In the Gallup Engagement Index Germany, only 21 percent of employees in AI-using organisations report that their manager actively supports and drives AI adoption. Double the span while expecting AI impact and you get neither.
Make decision logic explicit. This is the underrated part. Organisations document outcomes: processes, policies, handbooks. What they do not capture is the why behind them. Which option was rejected? How did anyone recognise that a case was an exception? Precisely that knowledge sits in people's heads, and precisely that knowledge is what both new managers and agents need in order to be more than fast and generic. At Omnora we have been working on exactly this point for years: not archiving experiential knowledge, but putting it to work.
The real question
The organisation of the coming years will be flatter. That has been decided.
Whether it also becomes smarter is decided by something else. Namely whether enough people remain inside it who can tell a good proposal from one that merely sounds good.
Breadth was a currency for a long time. You earned it over years, and it paid off.
Today you can get it on a monthly subscription. Twenty euros, without the years.
That is exactly how a currency is devalued. Whoever holds a lot of it has not become worse. They simply no longer get anything for it.
What is becoming scarce is the ability to understand something well enough to disagree with it.
Which decisions in your company can nobody assess on the merits any more, only pass along?
In part 2: why this question does not land in HR development but on the org chart.
Sources
- Gartner, press release of 22 Oct 2024: forecast that by 2026, 20 % of organisations will use AI to flatten their structure and eliminate more than half of existing middle management positions
- Stanford Digital Economy Lab, Brynjolfsson / Chandar / Chen, Canaries in the Coal Mine?, August 2026 version, ADP data through June 2026: 19 % relative employment gap among 22- to 25-year-olds in AI-exposed occupations
- Stepstone, analysis of 4.6 million job advertisements: share of advertised entry-level positions in 2025 around 42 % below the five-year average
- Randstad-ifo HR survey Q4/2025, published 19 Jan 2026, 500 to 1,000 HR decision-makers: 14 % today, 40 % planned within three years, 63 % among large companies
- Gallup Engagement Index Germany, published March 2026, 1,700 respondents: 21 % experience active support for AI use from their manager
- Gallup, State of the Global Workplace 2026: manager engagement down from 31 % (2022) to 22 % (2025)
- Stifterverband / McKinsey, AI skills in German companies, January 2025, around 1,005 executives surveyed
- Roland Berger, The AI-First Organization, 472 executives, fieldwork December 2025 to April 2026
- Kienbaum: rule of thumb of seven to ten direct reports per manager
- Business press 2025 and 2026: reduction of management positions at VW, BMW, Bayer, Deutsche Bahn, BASF, SAP, Evonik
