AI creates enormous opportunities to make work faster and better. At the same time, there is growing concern that we may lose skills as we delegate more and more work to AI.

But delegating work – and with it knowledge, routines and practical skills – is nothing new. Specialisation and division of labour have existed for a very long time, on a larger scale at least since the Bronze Age, around 5,000 years ago.

The same principle shapes the way we work today. Organisations routinely work with external partners, agencies, consultants and other specialists. In doing so, they also outsource knowledge and capabilities. Often, however, this involves specialist expertise that not everyone within an organisation needs to have themselves.

With AI, the situation is different. It is increasingly taking over precisely those fundamental tasks that have traditionally been part of our day-to-day work: research, drafting, comparison, summarising and review. And with them, tasks through which expertise has traditionally been built.

One common response is that we should continue doing at least some of this work ourselves so that we do not lose the underlying skills. I am only partly convinced. In an efficiency-driven working environment, it is questionable how realistic it is to deliberately retain tasks simply to preserve the capabilities associated with them. More importantly, this approach is strategically too limited.

The real question is which capabilities we need to retain despite delegating the work. I do not need to write every first draft myself, but I do need to recognise whether a text is good. I do not need to perform every analytical step manually, but I need to understand enough to challenge the result. Responsibility does not require us to perform every step ourselves. It requires judgement.

The importance of that judgement is particularly visible in AI-generated communication right now. Many LinkedIn posts sound increasingly alike, are roughly the same length, follow the same narrative patterns and end with the same engagement questions. I share the criticism. This uniformity does not just make the content less appealing to read. It also suggests lower-quality content, because it tends to emerge when AI output has not been sufficiently steered, reviewed and refined. What is missing is precisely what is needed to turn AI output into a sound, high-quality result.

This is where managing AI transformation carries a clear responsibility: not only to enable the use of the technology, but to shape new ways of working in a way that preserves quality and human judgement.

It is also too narrow to equate AI competence primarily with compliance or prompting. Both matter. But using AI competently requires more. What matters is whether someone can judge when AI is useful, which information or perspectives may be missing, how reliable an output is and when further intervention is needed.

AI can significantly improve the quality of complex knowledge work, not least because it gives us access to a much broader range of information and perspectives. But this does not happen automatically. Reliable quality depends on someone understanding the context, actively steering the AI, and assessing and correcting its outputs. The body of available knowledge may be vast, but it is neither automatically accurate nor automatically relevant.

And this is where, for me, the most difficult question begins: not how we can preserve every capability we have today, but how the next generation will build the expertise required to consistently produce good results with AI.

Much of what constitutes my expertise today comes from experience. Assessing situations, recognising patterns, formulating sensitive messages or judging quality are not things we learn in the abstract. That judgement has developed over years through exactly the kind of detailed foundational work that is now particularly easy to delegate to AI.

So what happens when these tasks increasingly disappear?

If organisations automate large parts of typical entry-level work, expertise cannot continue to develop in the same way. If research, drafting, comparison and review increasingly move to AI, development paths will have to change as well. Otherwise, organisations may use existing expertise more efficiently while simultaneously weakening the conditions under which new expertise can emerge.

Simply retaining parts of the old way of working on a smaller scale does not solve the problem. Most importantly, it does not ensure that early-career professionals gain the breadth and depth of experience from which expertise eventually develops.

A strategic response therefore needs to start somewhere else: Which capabilities will be needed in future? Through which experiences have early-career professionals developed those capabilities until now? Which of those learning environments are disappearing because of AI? And which new experiences could build expertise that is more relevant for the future?

This gives AI implementation a sustainability dimension.

Managing this process should therefore consider not only adoption, productivity and governance, but also which human capabilities an organisation will need in the long term and how they can be developed in future. Critical expertise cannot be assumed to reproduce itself when the conditions under which it has traditionally developed are disappearing. In my view, AI implementation should therefore also include a capability sustainability perspective.

This question extends far beyond individual companies. If many organisations automate the same developmentally important tasks, individual productivity decisions become a broader societal question: where will future expertise come from?

Perhaps this is also a question for ESG – and possibly one that has already been asked: whether responsibility for preserving human capability in the age of AI should be anchored more clearly and explicitly within its social dimension.

The challenge is not to prevent work from being delegated to AI. It is to ensure that people will still be able to develop the expertise required to use AI effectively and assess its outputs. Without that expertise, the quality gains AI can enable cannot be sustained in the long term.