A paradigm shift: AI isn’t targeting manual work. It’s targeting your highest-paid experts

Manual work isn’t safe; it will simply be affected later. The first to feel the impact will be the people companies pay the most, and the budgets used to develop them.

Ask your leadership team where automation will happen first. You’ll hear warehousing, manufacturing or invoicing. The places where the lowest-paid routine work gets done.

That’s how every previous wave of automation worked. From the steam engine to industrial robots, automation targeted manual, repetitive tasks first, and later simple administrative work. A company bought a machine or system and only then decided what would happen to the people.

Now it’s different.

In this article, I’ll look at three things happening at the same time. The data shows that AI already affects a large share of the work done by well-paid specialists. Employees are adopting it faster than companies can govern it. And the value of expertise is shifting from producing output to verifying it. All three lead to the same question: how should companies plan work and employee development in this situation?

The highest-paid workers are the most exposed

Anthropic (the developer of Claude AI) published an analysis in March that measures which work tasks AI is actually performing rather than predicting what it might do. Computer programmers top the list, with such tasks accounting for 74.5% of their working time.

You won’t find warehouse work there. Manufacturing or invoicing either. The entire top half of the ranking is made up of experts companies pay a premium for.

Exposure does not mean layoffs. The study measures which tasks the model contributes to, not how many jobs disappeared because of it or how many junior employees companies chose not to hire.

Share of working time made up of tasks the model contributes to in practice. Based on Claude usage in the second half of 2025. Source: Anthropic, Labor market impacts of AI, March 2026.

There is one caveat to the table. The figures come from usage data for a single tool (Claude AI), which is widely used for programming, so programmers also rank highly because of who uses Claude. Occupations whose tasks barely appear in that usage data show zero in the table.

I don’t see a reason to think a fundamentally different group of occupations would emerge with other AI tools. Similar people use them for similar tasks. The order at the top might shift, but not which occupations appear in the upper part of the table.

The same Anthropic analysis looks at one more thing: how much people in different occupations earn. The most exposed quartile earned 47% more than occupations that did not appear in the data at all.

The wage data is from autumn 2022, just before ChatGPT was released. Wages have moved over the past four years, but that does not change the core point. These were already some of the highest-paid occupations before generative AI entered the commercial market.

We are used to the opposite from automation. In the past, it made simpler work cheaper and increased the value of educated workers because they operated the machines that displaced others. Now the model can handle precisely the work that used to command a premium.

Employees adopted AI themselves

In the past, automation could enter a company only through a purchase. A production line, crane or accounting system went through investment planning and approval, so the company knew in advance what it was buying and when it would arrive. At the same time, it could plan how the change would affect employees. Someone had to decide who would operate the new machine, what they needed to learn and how the company would know whether they were doing the job well.

AI has no such time or process barrier. It can be deployed through a browser, paid for monthly and often requires no operational rebuild. Employees therefore have little reason to wait for the company to decide. They open a tool that removes the parts of their job that slow them down most, without asking anyone.

Nearly nine in ten companies have invested in AI, yet fewer than four in ten report measurable impact. Three quarters of people in knowledge-work roles already use it at work. Many of them use tools they found themselves and that their companies have never formally rolled out.

Work is shifting from producing output to verifying it

Because employees adopted AI themselves, the point where companies need expertise has shifted too.

But no one gave them a brief defining what a good output should look like. So everyone decided for themselves what was good enough, based on what they could get accepted. The nature of the work changed, and with it the standard companies used to judge whether someone was doing the job well.

Judging the output and saying it is wrong still falls to a human. But only someone who could do the work themselves can do that well.

Companies can ban AI, but that solves nothing. They will not get it out of their employees’ work and will only lose visibility into how people are using it. They will only slow themselves down.

Employee development must be planned as an investment

Companies know the answer. Training. According to the World Economic Forum, 77% of companies plan to launch upskilling or reskilling programmes. The OECD also tracks how many adults actually take part in further education, and in many countries that figure is flat or falling.

But that only tells us how many people attended a course. Few companies check whether anything from it carried over into how they work. Knowledge becomes experience only when it is applied on the job — and only then can the money invested start to show a return.

The people responsible for AI inside companies see the same problem. McKinsey surveyed around 500 organisations about what prevents them from using AI safely and with confidence. Nearly 60% of those responsible for AI and its risks said employees lacked the knowledge and training to use it properly. A year earlier, the figure was around half.

I think the reason is that companies do not manage employee development as an investment. Training budgets are set, often based on last year’s figure, and then have to be spent. So companies train people on all sorts of things simply to make sure nothing is left in the budget at year-end (and to avoid having the budget cut the following year). Priority goes to whoever asks for a course. Training hours are counted.

Yet no one asks where the content of the work has changed over the past year. A company that cannot answer that question does not have a development plan. It has a budget and a catalogue of courses.

When work becomes about verifying someone else’s output, two skills in particular need developing. People must understand what the technology can do and where its limits are. And they must be able to challenge an answer that sounds convincing.

The text looks finished even when the AI model has made up some of the information. The only way to catch this is to verify where the information came from and check whether the output matches reality. The second skill is critical thinking, and it is just as important. With both in place, uncertainty becomes a concrete development need that a company can actually act on.

What this means for your development plan

AI is not taking experts away from companies yet. It has taken away the certainty that a high-quality output was created by the expertise of the person who submitted it. A development plan therefore now depends on whether the company knows which roles are changing, who checks AI-generated outputs, and whether those people have enough knowledge, and enough room, to say when the output is wrong.

Next time you approve the development budget, ask two questions. Do we know which roles have changed over the past year? And how much of that budget is going to those roles?