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.
In every organisation I know, there were managers who did not carry their area on the merits.
Not lazy. Not malicious. Usually pleasant, often loyal, sometimes with twenty years in the building. Just without a professional foundation in the field they were responsible for.
And it worked anyway. There was a strong deputy. There was a team that quietly filled the gap. There were processes that caught mistakes before they cost money.
That was expensive. But it was bearable.
That buffer has just disappeared. And not slightly.
What has changed arithmetically
AI does not change a person's error rate. It changes their throughput.
In the past a manager might make three or four substantial decisions a week. Each one had lead time. There was a paper, questions, a committee, time. Time in which it could become apparent that the thing had not been thought through.
Today, at the same desk, there are dozens of fully formulated results per day. Argued, with numbers, in good prose, in the right structure.
And the only remaining task at that interface is review.
Which is exactly the capability this person lacks.
The point is not that they decide wrongly. The point is that they can no longer not decide. Every wave-through is a decision. Approve fifty papers a day without being able to assess one of them and you have made fifty decisions you did not make.
The finding that makes this uncomfortably concrete
The research lab METR ran a randomised controlled trial in 2025. Sixteen experienced open-source developers, 246 real tasks, in codebases they had known for an average of five years.
Beforehand they expected to be around 24 percent faster with AI. Afterwards they believed they had been about 20 percent faster.
Measured, they took 19 percent longer.
That number became famous. The genuinely important finding sits next to it: the developers accepted less than 44 percent of the code suggested to them. They threw away more than half. And they spent around 9 percent of their working time reviewing and correcting the output.
That is judgement in operation. Visible as a cost centre.
Now the question this piece is about.
What happens when the person at that desk cannot spot the discarded 56 percent?
They are fast. They feel productive. They deliver on time. And nobody notices anything until it gets expensive.
Note the second part of the METR finding while you are at it: even experts misjudged their own speed by nearly 40 percentage points. Those with depth are wrong about themselves. Those without it do not even have the material to be wrong with.
Why this escalates in leadership roles
A caseworker who accepts poor AI output produces a faulty case.
A manager who does the same produces a standard.
They approve. They prioritise. They do not escalate. Their judgement is the last instance at which a plausible error might still be caught before it goes outside or migrates into a system. Remove that instance and the error becomes a decision, and the decision becomes a precedent.
There is a second effect that works more slowly and echoes longer: you cannot teach what you cannot assess. A team learns from its manager what counts as good enough. If the answer is anything that looks plausible, that is what it learns.
Two years later you have a department in which nobody can say no.
Slowness was a safeguard
In part two I described the classical organisation as a rationing machine: briefing, specification, approval, backlog. All mechanisms for allocating scarce delivery capacity. And all friction, once that capacity is no longer scarce.
That still holds. But it is only half the description.
Those loops were not only rationing. They were also redundancy. Four pairs of eyes at four points. If three missed something, the fourth caught it.
Cut layers while multiplying throughput and you remove both at once: the friction and the safeguard.
This is why restructuring without building competence is not neutral. It is risky. And this is also why the opposite is no way out: stay slow to stay safe and you lose the safety anyway, because the people in the chain remain the same.
The uncomfortable part
Keeping a manager without professional depth was long a social decision. It had a price, and the price was paid in efficiency.
Today it is a risk decision. The price is paid in quality, liability and reputation, and it does not fall due in the quarter the decision was made.
This is explicitly not a call to weed people out. Most of these people are not unsuited. They are misdeployed or were never developed, often by exactly the organisation that is now disappointed. Someone promoted for coordination for twenty years has not neglected professional depth. They did what was rewarded.
But a decision has to be made, and it has to be said out loud. Either build depth, with time, budget and a realistic deadline. Or change the role without damaging the person.
What no longer works is the third variant, the one currently running in most companies: do nothing and hope the organisation absorbs it again.
It no longer absorbs it. That was precisely the service of the layer being cut.
Three things you can check in the coming weeks
Build the list that exists nowhere. At which points in the company does an approval currently sit with someone who cannot assess the content on the merits? That list exists in no org chart and in no risk register. Building it takes two weeks and changes every subsequent discussion about AI.
Measure the rejection rate, not the adoption rate. Almost every company today reports how many employees use AI. Almost none reports how often AI output is rejected. At METR that rate was over 56 percent, among experts, on familiar ground. If it sits near zero in your company, that is not a quality signal. It is a warning signal.
Make the safeguard explicit. Which decisions absolutely require a second, professionally competent pair of eyes, and which do not? The answer appears in no process manual, because it never had to be said. It sits in the heads of the experienced. Translating exactly that knowledge out of heads into something usable is the work we have been doing at Omnora for years. Without it nobody reviews anything, human or agent, because nobody knows how to recognise that something does not hold.
To close
There have always been people in leadership roles who did not carry their area on the merits. The organisation carried them, because it was slow enough to afford it.
That indulgence was never a virtue. It was a by-product of inertia.
The inertia is gone.
Where in your company does an approval today hang on someone who cannot assess the content?
Sources
- METR, randomised controlled trial, first half of 2025: 16 experienced open-source developers, 246 real tasks in their own repositories. Expectation beforehand plus 24 % speed, self-assessment afterwards plus 20 %, measured 19 % longer completion time. Less than 44 % of suggested code accepted, around 9 % of working time spent reviewing and correcting output. METR itself describes the result as a snapshot of early 2025 and warns against generalisation.
- MIT Project NANDA, The GenAI Divide: State of AI in Business 2025: the learning gap in tools and organisations as the main cause of absent bottom-line effects.
- Gallup, State of the Global Workplace 2026 and Global Indicator: Artificial Intelligence: the discrepancy between individual productivity gains and organisational impact.
- Roland Berger, The AI-First Organization, fieldwork December 2025 to April 2026, 472 executives: 59 % consider their own leadership insufficiently prepared, 49 % name missing AI capabilities as the biggest hurdle.
