AI at work: what happens to technical expertise?
The relationship between technical expertise and AI is becoming an increasingly important question for employers.
In just one year, the share of Swiss SMEs using artificial intelligence in their business processes has risen from 22% to 34%.
Among companies already using AI, one in three says the technology is changing the skills they look for when recruiting.
This raises a question that receives far less attention:
When a tool does part of the work for us, how do we maintain the skills we need to judge whether its output is good?
Recent research is beginning to provide some answers.
When AI changes the nature of work
Our previous article highlighted a finding that may seem surprising.
The rise of transversal skills in Swiss job advertisements began before the arrival of generative AI. These skills are also growing at roughly the same rate in occupations that are highly exposed to AI and those that are less exposed.
AI, therefore, does not explain this shift.
But another question remains.
What happens to technical expertise when AI becomes part of everyday work?
What research is starting to show
Researchers from Microsoft Research and Carnegie Mellon University surveyed 319 professionals about their use of generative AI. They analysed 936 examples of real workplace tasks. Their study was presented at the CHI 2025 conference.
The researchers found a change not simply in how much people worked, but in what kind of work they did.
- Searching for information became verifying information.
- Solving a problem became integrating an answer that had already been produced
- Execution became supervision.
In other words, the work did not disappear. It changed.
But producing something and checking it require different cognitive skills. And we do not necessarily maintain those skills in the same way.
The researchers also found an important relationship between confidence and critical thinking:
The more people trusted the AI tool, the less critical thinking they tended to apply. The more they trusted their own expertise, the more likely they were to question the AI’s output.
For organisations, this matters.
Knowing how to use an AI tool does not automatically give someone the ability to identify an error in its output.
That ability comes largely from professional expertise developed through experience and practice.
An experienced employee may quickly recognise that something is wrong. Someone at the beginning of their career may find that much harder.
A second study, from the MIT Media Lab, approached the question differently.
Researchers followed 54 people completing a writing task. They divided participants into three groups: one used an AI assistant, another used a search engine, and the third worked without external assistance.
Using electroencephalography to measure brain activity, the study by Kosmyna and colleagues recorded the lowest level of brain activity among participants using AI assistance. Participants working without external assistance showed the highest levels.
The study has attracted scientific debate.
A commentary published in early 2026 highlighted the small sample size and urged caution when interpreting the results.
That caution is important.
But both studies point towards the same broader question:
If regular practice helps us maintain our ability to judge the quality of work, what happens when we practise those skills less often?
This is where training and the organisation of work itself becomes important.
We previously described this mechanism as Skill Erosion, alongside Shadow AI and AI Washing.
Since then, researchers have started to build a stronger evidence base around the phenomenon.
The limits of the “we’ll adapt” argument
Several arguments tend to come up when organisations discuss this issue.
One is that companies have adapted to new technologies before. The comparison with tools such as spreadsheets makes sense up to a point. But the technologies do not work in the same way.
A spreadsheet performs an operation requested by the user. Generative AI can propose the answer itself. That difference matters because it can shift the employee’s role from producing work to assessing work.
Research is only beginning to measure the effects of that shift, but the first findings suggest that it deserves attention.
Another argument is that experienced employees already know how to check the results. In many cases, they do. The available evidence suggests that professional expertise helps people assess AI-generated output more critically.
But this creates another problem:
How will junior employees develop that expertise if AI already performs many of the tasks through which previous generations learned?
Someone cannot learn to recognise a weak argument, an incorrect calculation or a poor recommendation purely by being told what good work looks like.
Expertise also develops through practice.
A third argument is that organisations should wait for more evidence.
This is also a reasonable position. Research on generative AI at work is still young, and many findings remain open to debate. But waiting still requires observation. If output remains high and deadlines are met, skill erosion may be difficult to see.
Organisations therefore need to ask:
What would tell us that a team’s underlying expertise is weakening?
Without indicators, companies may not notice the change until that expertise is needed.
What Swiss companies are already seeing
Swiss data give us another view of the shift.
According to AXA’s study of the SME labour market, conducted by the Sotomo research institute among 300 companies in French- and German-speaking Switzerland, AI adoption among SMEs rose from 22% to 34% in one year.
But another figure may be even more significant.
One third of SMEs using AI say the technology is already changing the profiles they look for when recruiting.
Michael Hermann, Director of the Sotomo Institute, notes that companies increasingly want employees who are comfortable with technology and willing to keep learning throughout their careers.
These two capabilities require different approaches.
A company can often teach a specific technological skill through a course or training module.
The ability and willingness to keep learning develop differently. They depend on team habits, management practices and organisational culture.
This makes AI adoption a management issue as much as a training issue.
Training is only part of the answer
Switzerland does not start from zero.
According to the Federal Statistical Office, almost half of the working population participated in job-related continuing education in 2021, the latest year for which detailed survey data are available.
Of those who participated, 93% received support from their employer.
Switzerland therefore already has a strong culture of continuing professional development.
AI training, however, is still taking shape.
A Deloitte Switzerland survey, conducted in autumn 2025 among 99 executives, found that AI training was voluntary in 55% of organisations and mandatory in 24%.
The sample is relatively small, so these figures should be interpreted with care.
But they reinforce a broader point: many organisations are still deciding how AI training should fit into their learning strategy.
The challenge is not simply to teach employees how to use the tools.
It is also to maintain the judgement required to challenge them.
What should organisations preserve?
Before allocating a training budget, companies can start by asking a few practical questions.
Training the tool is not enough
Teaching employees how to use an AI assistant solves an immediate problem. But that knowledge may have a short lifespan. AI tools change quickly, and training must constantly keep pace.
Teaching employees how to assess an answer serves a different purpose.
- Can they identify what is missing?
- Can they recognise a weak argument?
- Can they detect unreliable information?
- Can they decide when not to trust the output?
These skills remain useful even when the tool changes.
Organisations need both.
The real question is how to balance them.
Who still knows when the output is wrong?
For every process that uses AI, ask one simple question:
Who can currently recognise when the tool gets something wrong?
If someone has the expertise but is not responsible for approving the final result, you may only need to change the process.
If nobody can reliably identify the error, you have identified a skills gap.
That gap requires training, practice or access to additional expertise.
Some skills should not be outsourced
Look at tasks your team performed directly two years ago that are now partly or fully delegated to AI.
Then ask:
Is it acceptable to stop practising this skill?
For some tasks, the answer may be yes.
Translating a routine administrative letter, for example, may not require employees to maintain the same level of manual practice.
Building an argument for an important client is different.
So is interpreting complex data, making a strategic recommendation or assessing a sensitive situation.
There is no universal answer.
The value of the exercise lies in making the decision deliberately rather than by default.
Building expertise before supervising AI
The question becomes especially important for people entering the workforce.
If employees need expertise to assess an AI-generated result, junior professionals face a paradox.
They may benefit greatly from AI because it helps them work faster.
At the same time, they have had less time to develop the expertise they need to recognise when the AI is wrong.
Organisations therefore need to think carefully about which tasks junior employees should still perform themselves.
This does not mean keeping AI away from them.
It means designing opportunities for them to build expertise before asking them to supervise it.
Training may also need to differ according to experience level:
- A senior professional may need to learn how to integrate AI efficiently into an existing workflow.
- A junior professional may first need opportunities to understand the underlying task.
Keeping human expertise at the centre
AI does not make technical expertise irrelevant.
It changes how organisations need to maintain it.
Research increasingly suggests that generative AI moves part of the employee’s role from execution towards supervision.
But good supervision depends on judgement. And judgement depends, at least partly, on knowledge and experience built through practice.
Swiss companies are already seeing the consequences. One third of SMEs using AI say it is changing the skills they seek when recruiting.
For HR and L&D teams, this creates a new balancing act:
How much should organisations invest in teaching employees to use AI and how much should they invest in maintaining the expertise needed to challenge it?
That question also leads directly to the final article in this series.
If people develop and maintain skills through practice as much as through formal training:
How do adults actually learn?
Want to develop these skills within your team?
Swissnova offers training designed to build skills that last. Contact our team for more information tailored to your organisation’s meeds.
