A piece for boards, managing directors and C-level.
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.
Part one was about the ability to tell correct from plausible, and why it grows out of real depth. That sounds like a topic for HR development.
It is not. It is a structural topic, and it has a deadline.
The most quoted figure of the past year comes from MIT: around 95 percent of the generative AI pilots examined delivered no measurable effect on the profit and loss statement.
The figure is contested, and rightly so. The report from MIT's Project NANDA is preliminary, not peer-reviewed, rests on a small interview base and defines success narrowly through measurable bottom-line effects after roughly six months. The criticism lands on the precision of the number. It does not land on its direction.
Because the interesting finding is not the number. It is the reason. The projects did not fail on the models. They failed because afterwards not a single process ran differently from before. MIT calls it the learning gap.
The Roland Berger study The AI-First Organization, for which 472 executives worldwide were surveyed between December 2025 and April 2026, makes the same picture visible from another angle: 62 percent expect major or radical changes to their operating model. Only 38 percent have started the transformation at all. 59 percent say openly that their leadership is not sufficiently prepared for it. And the biggest hurdle named is not technology but missing AI capabilities, cited by 49 percent.
So between expectation and execution sits a gap of 24 percentage points. That gap has an address.
Who really decides whether AI lands
Gallup measured something that gets entirely lost in the debate about data centres and model sizes: US employees in AI-adopting organisations who credit their manager with actively supporting AI use are 8.7 times more likely to agree strongly that AI has changed the way work is done in their organisation.
8.7 times. Not 20 percent better. Almost ninefold.
In the Gallup Engagement Index Germany, only 21 percent of employees in AI-using organisations report that their manager actively supports and drives adoption. In the US it is less than a third.
And the second figure fits alongside it: 65 percent of US employees in AI-using organisations report personal productivity gains. But only 12 percent say AI has actually changed the way their organisation works.
Middle management sits precisely inside that gap.
The effect is multiplicative, not additive. A manager without AI competence does not simply forgo their own productivity gain. They neutralise the gain of their entire team. What they do not understand, they cannot prioritise, cannot defend and, when in doubt, will not approve.
Put differently: middle management is the amplifier of every AI rollout. In German companies, that amplifier is switched on for roughly one fifth of teams.
Why this is not a character problem
It would be convenient to talk about resistance to change here. That misses the point.
The value of middle management consisted for decades of coordination: condensing information, translating between layers, allocating capacity, reporting status. That was a real, scarce and well-paid service.
Now the same people are being asked to accelerate the adoption of exactly the technology that renders that service redundant. Without anyone telling them first what their new service is supposed to be.
Do that, then complain about a lack of willingness to change, and you have a design problem that you are calling an attitude problem.
So the question is not how to motivate this layer. The question is what you are paying it for from tomorrow.
The answer is builders
Not programmers. Builders.
By that I mean the ability to close the distance between a problem and a working thing yourself. Without a project. Without a requirements document. Without the detour through three departments and a quarter.
Why this changes everything structurally:
The classical organisation is a rationing machine. Briefing, specification, prioritisation, approval, backlog, roadmap: every one of those steps exists because delivery capacity was scarce and expensive. It had to be allocated, so you needed a layer that allocates.
That resource is no longer scarce.
A rationing machine for a good available in abundance creates no value. It creates only friction. And from the inside, friction feels like diligence.
This is why the equation fewer layers equals more speed falls short. Speed does not come from cutting approvers. It comes from the people who have an idea also being able to build it.
A manager who does not master the new tools themselves is not slow in this logic. They are a bottleneck for everyone who would be faster.
Restructuring fast is not cutting fast
This is the most expensive error of the year.
Many companies are currently cutting layers and calling it transformation. It is a cost measure. It shows up immediately in the accounts and changes nothing about how work gets done, because the remaining managers perform exactly the same role as before, only for twice as many people.
The result is the worst of all variants: less coordination capacity and still no new competence.
Restructuring means something else. It means changing the shape of roles before touching positions. Which decisions are actually made here? Which of them can an agent prepare? Which must it never make? What remains, and what does the person doing it need to be able to do?
That work is less comfortable than a cut list. But it is the only kind that still holds after two years.
Why speed genuinely matters here
I am generally wary of urgency rhetoric. On this point I consider it justified, for two reasons.
Structural debt compounds. Every quarter an organisation continues working in the old structure, it trains its people in the old behaviour. Processes set, expectations harden, career paths reflect what was rewarded yesterday. The restructuring does not get linearly more expensive. It gets harder.
The actual learning curve cannot be bought. Consultants and vendors supply models, tools and reference architectures. What they cannot supply is the knowledge of where AI reliably fails in your business, with your cases and your exceptions. That knowledge only comes from doing it yourself, and it comes slowly. Start today and in two years you have something that cannot be purchased later.
Everything else about AI can be bought. Not that.
Four consequences for the next two quarters
Evidence instead of training. AI training for managers produces attendance rates, not competence. The harder requirement: every manager has personally automated or rebuilt at least one real process in their own area of responsibility in the last quarter. Personally. Not commissioned.
Measure leadership by what got built, not by team size. As long as advancement depends on the number of reporting lines, everyone rationally optimises for more people rather than more impact. Some companies have now broken that link. This is not symbolic politics; it is the lever that turns behaviour most directly.
Capacity to restructure instead of a restructuring. The one big reorganisation is the wrong format. What you need is an organisation that reviews its own shape regularly, because the line between what people do and what systems do shifts every few months. Restructure every five years and you always restructure at the wrong moment.
Provide context before you demand speed. A builder without access to their company's decision logic builds quickly in the wrong direction. Speed without judgement is just being wrong faster. MIT names exactly this as the core of the learning gap: tools that do not retain an organisation's context and do not adapt to it. At Omnora we work on this point, because the experiential knowledge that makes agents capable of acting sits in people's heads and not in folders.
The question I would put to a board
Not: how large is your AI budget?
Not: how many pilots are running?
But:
How many of your managers have personally built something in the last 90 days that someone uses today?
That number is your actual AI strategy. Everything else is a statement of intent.
In part 3: why missing judgement in leadership roles no longer merely slows things down but does damage.
Sources
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025: 95 % of the GenAI pilots examined without measurable bottom-line effect; the report attributes this to the learning gap in tools and organisations. Preliminary report, not peer-reviewed, sample figures vary across accounts
- Roland Berger, The AI-First Organization – Turning AI power into enterprise performance, published 6 July 2026. 472 executives, fieldwork December 2025 to April 2026, 35 % from the DACH region: 62 % expect major or radical changes to the operating model, 38 % have begun, 59 % consider leadership insufficiently prepared, 42 % doubt their governance structures, 49 % name missing AI capabilities as the biggest hurdle
- Gallup, Global Indicator: Artificial Intelligence and State of the Global Workplace 2026, based on a US workforce survey in Q1/2026: factor of 8.7 for strong agreement that AI has changed how work is done in the organisation; 65 % personal productivity gains against 12 % organisational change
- Gallup Engagement Index Germany, published March 2026, 1,700 representative respondents: 21 % experience active support for AI use from their manager
- Stifterverband / McKinsey, AI skills in German companies, January 2025, around 1,005 executives: 79 % say their employees lack basic AI skills
