In the article I shared last week, I wrote that the task of rolling AI out across the organization responsibly belongs to HR. I still think the same. I believe this is not a software installation matter but a people and culture matter, and that the function taking on this task must first become fluent itself.
Now I think it is important to show you the other side of the coin as well.
As AI adoption spreads, two things happen at the same time. On one side, the need for headcount is falling; we are already discussing that side with numbers, workforce plans and efficiency calculations. On the other side, the skills of the people who are working do not stay at the same level but start trending downwards. Frankly, the second one has neither a report nor an indicator.
The question I am looking to answer in this piece is: "With AI, which skill exactly is eroding, how is it eroding, and what can we do about it?"
Let's continue with an old technology case.
An accounting firm in Finland decided to remove the automation system it had used for years from its architecture. When the system was switched off, the firm saw that its accountants could no longer do the core work of that process. A research team from Aalto University studied this case over several years and published its findings in the academic journal of the Association for Information Systems.
The system in this case was not artificial intelligence in today's sense; it was rule-based automation software doing a limited job. The reason I bring the case here is to convey that skill erosion did not begin with AI; it is an old mechanism that operates everywhere we hand work over to technology. What makes AI different is that the same mechanism now runs across a far wider set of tasks, much more intelligently and much faster. In that firm, what eroded was the skill of a single process. What we are talking about today is nearly all work done by thinking.
And there is this: the loss was not clearly noticed while it was happening – it cannot be noticed. It surfaced once the system disappeared. I will come back to this detail at the end of the article.
Which skill is eroding?
Let's first describe what erodes, because these concepts can be confused.
First of all; what erodes is not "knowledge". We have long kept knowledge in the tools and devices we use; in books, files, phones. Not remembering the formula of something has not been anyone's problem for a while.
What erodes is "the kind of skill we gain while doing the work with our own hands". To make the concept clearer, we can define it as "being able to frame a problem correctly, reach a decision with incomplete information, test an assumption, build a synthesis from scattered pieces and understand whether an output placed in front of you is good or bad".
What these have in common is that they cannot be learned by being told. No one can teach you judgement as a lesson. Judgement is built like this: you make a decision with incomplete information, you see the outcome, you update the model in your mind and next time you decide a little more accurately. That is, through repetition and struggle.
Harold Jarche describes these skills as hand-made work and says:
"Sensemaking is a hand-made skill; it can be supported by various tools, but the most important tool is our mind. Hand-made skills are human skills."
BCG's study with 70 senior executives arrives at the same list. The competency carrying the highest erosion risk is judgement and decision-making; the competency seen as most critical for corporate performance is understanding and framing the problem. We can say that the skills the company needs most and the skills eroding fastest are in the same basket.
By contrast, empathy, active listening, curiosity and leadership impact stay in a relatively protected zone. In other words, skills grounded in human relationships and bodily experience remain standing.
So how does it erode?
Jarche's rule is simple: if you don't use it, you lose it. "Automate a process that was once done with skill, and that skill declines."
As for how the mechanism works, the skills I listed above were actively used while doing the work in person. While building an analysis from beginning to end yourself, while producing the skeleton of a report from scratch, while breaking a problem into pieces. Today, these are exactly the first tasks we hand over to AI. In other words, the most easily delegated work and the work that builds the skill are the same.
When you do an analysis yourself you get bored, you get stuck, you turn back from wrong paths, and this difficulty looks like inefficiency from the outside. Yet that is where learning happens. Ross Stevenson, who writes in the L&D field, puts it this way in his article on skill erosion:
"People largely want to take the path of least resistance and choose instant gratification rather than struggle. The problem is that real learning, and therefore improving performance, is about going through difficulty. Without experiencing struggle it is genuinely very hard to learn the how, the what and the why."
The invisible trap here is that the process feels like learning. While working with AI you see a good piece of work and output on the screen, you read it and approve it. Your mind records this as a success. The warning Stevenson placed in his article a year earlier is exactly this: it is very easy to convince ourselves that we have learned to do something, whereas the one doing the work is the AI.
The second warning in the same article also matters: AI does not take these skills on its own — the way we use it is what actually shapes it. To make this concrete; two different things can be done with the same tool. In the first, you tell the tool "write this report", you read and correct what comes out; in this usage, framing, synthesis and decision steps all stay with the tool. In the second, you first build your own draft and argument, then tell the tool "strengthen this, show me its weak spots"; in this usage those steps stay with you and the tool only speeds you up. Same tool, same time, two different outcomes. I would say the difference is hidden in how the tool is used.
The second mechanism is trust in these tools slowly increasing. The tendency Yulin Fang of the University of Hong Kong Business School describes as automation bias: as the quality of the output rises, the user spends less and less effort verifying it. The Aalto team's model also shows the organizational counterpart; growing trust creates complacency and a comfort zone, complacency loosens the mechanisms that keep the skill alive, and the loop repeats itself.
Third is the linearity of the usage pattern. Describe the task, get the answer, fix the output. According to BCG's finding, this pattern completely bypasses the cognitive step that builds judgement. One interviewed executive puts it clearly:
"Teams start treating AI output as 'good enough' and skip the critical thinking step entirely. Over time no one looks for gaps in the analysis; errors and blind spots go unnoticed."
The fourth, and perhaps the most hidden and insidious mechanism, is that the loss appears with a delay. The point Esko Penttinen of Aalto underlines in his interview: as long as the system works there appears to be no problem, because the work gets done and the output arrives. The eroded skill becomes visible only when the system goes out of service or an unusual situation arises. At that moment it becomes clear that people cannot do the work on their own. That is exactly what happened in the Finnish case.
Another sentence of Penttinen's from the same interview shows the weak link in the efficiency argument. Automation's classic promise: let's give the boring work to the machine so people can focus on higher value-added thinking. Penttinen points out where this promise falls short; that value-added thinking cannot be acquired without going into the detail. While the accountant deals with invoice data they become immersed in the fine details of the process, and insight is in fact formed inside that immersion. In other words, within a process and a skill whose details you do not master, you may not create value either.
But what if these skills really are becoming unnecessary?
Up to here I have written from a single point of view. But examining the other side's perspective will be important to support the research and the approach I put forward, because it deserves to be taken seriously.
Similar things were said when the calculator became widespread. Today no one mourns the loss of the ability to do arithmetic by hand. Navigation arrived and map reading died; few think there was a cost to that. Every generation dropped a skill and the world did not become a worse place. Maybe judgement will go the same way; maybe in ten years saying "build your own analysis" will sound as odd as saying "use a slide rule yourself" does today.
I do not dismiss this possibility. But I think it differs on three points.
Previous tools took over a single job and their boundary was clear. The calculator does the multiplication, you decided what to multiply by what. Today's tool can take that decision step as well; what you delegate stops being an operation and becomes judgement itself.
With previous tools you held the ground to check the output. The calculator would not give a wrong result. AI models can, and noticing that error requires precisely the skill that is eroding. This looks like a closed loop, doesn't it?
What is needed to check is eroding because you are not checking.
The third is Penttinen's own question.
"Which skill should be preserved, which can be let go?" It is hard to draw this line today, because a competency that looks unnecessary now can become critical when the environment changes. With the calculator the stake was low; here the stake is the company's decision quality.
I should also say this: the evidence in this field is still early-stage evidence, no one has ten years of tracking data. What we have is one case study, one field experiment, one laboratory study and one executive survey. Their methods are independent of each other and they all point in the same direction. Not enough for a definitive verdict, but more than enough for a risk assessment.
The experiment conducted in a high school in Türkiye
Everything I have described so far has been measured by a single experiment, and I find it additionally notable that the experiment was conducted in Türkiye.
A research team from the University of Pennsylvania and The Wharton School ran a controlled experiment with about a thousand mathematics students in a high school in Türkiye. We have 3 groups:
The first group used an interface called GPT Base, similar to ChatGPT; this interface gives the solution steps and the answer directly. The second group used an interface called GPT Tutor, constrained by teacher input; this interface gives only a hint at each critical step, but does not give the answer. The third group is the control group; it used no AI tool at all, working only with the textbook and notes.
In the practice session the hint group performed 127% better than the control group and the answer group 48% better. Up to here both groups look like they gained.
In the other part of the research, where the striking results appear, in the exam taken without AI the answer group performed 17% worse than the control group that had never used AI. The hint group stayed at the same level as the control group in the exam; that is, the negative effect disappeared.
I would like to dwell on these results a little. Same school, same course, similar students and the same underlying model. The only thing that changed is whether the tool gives the answer. The group that got the answer looks good during practice and, when left alone, ended up worse off than having never used AI at all.
Another finding of the research confirms the trap I described above: students using AI were overly optimistic about their own level of learning. So they were not only learning less; they thought they were learning more.
Think of the organizational counterpart of this. Two analysts in the same team. One has the model write the report and reviews it, the other first builds their own draft and then improves it with the model. At the end of that quarter we can say there is no difference between their outputs; both delivered the work on time and properly. But while we will see a serious gap between them 2 years later, no performance tool will show that gap. Because the performance form looks at the output, and the output has stayed the same since day one.
Where does accountability sit?
This sentence from the summary of the Aalto research sits, I believe, at the centre of the discussion:
"Accepting the erosion of skills that have become unnecessary is required to reap the benefits of technology. But the erosion of core human expertise is problematic if employees continue to be held responsible for work they do not sufficiently understand; they are left unable to respond when the automation fails."
The balance in this sentence matters. Work is delegated, accountability is not. Your name is still under the output; your ability to produce it is slowly eroding.
Stevenson asks the same question from the individual side:
"But what if the reason you lose is not that AI does your job better than you? What if you lose because you leaned too much on the temporary power it gave you? Then you become the architect of your own downfall."
Nobody wakes up one morning and decides to delegate their judgement. Having an AI tool write a draft, letting an analysis pass without verification, having the model build the flow of a presentation, and so on… As Stevenson reminds us, life consists of a series of trade-offs; when you say yes to one thing you have said no to another. Looked at one by one, none of them looks like a decision. Repeated hundreds of times, they set the direction of a career.
The observation of an executive BCG interviewed describes well how accountability shifts:
"I can see employees' sense of ownership declining. If the outcome is not positive, responsibility is placed on the AI."
So what is the individual counterpart of this? I think we need to pay attention to three places:
First, the moment work goes outside the usual. While the tool gives adequate results in standard situations, in an unexpected situation no one can give the answer and it is your name that remains as the responsible party.
Second, as the career progresses. As seniority increases the work expected of you changes; rather than producing output yourself, you evaluate whether the output your team produces is correct. And to be able to tell whether work has been done correctly, you need to have done that work with your own hands at some point. Without that experience, as a manager you cannot be sure what you are approving.
Third, in market value. The analogy Stevenson uses becomes very clear here: AI gives you an exoskeleton, and when you put it on you do work you normally could not — except everyone has the same exoskeleton, and when you take it off you are left with your own body. In a market where everyone accesses the same tool, what distinguishes you will be the judgement and experience the tool cannot produce.
What the three have in common is that the price is paid much later. What you gain today by delegating work to the tool is concrete and immediate: a few hours of time. What you lose is abstract experience that accumulates very slowly. The distance between them can be measured in years, and during that time you receive no warning. And when you notice, making up the loss will take longer than learning from scratch.
Let's first narrow the scope
Reading up to here, a wrong inference can be drawn: it may be understood as be restrictive with AI or slow down, but my aim is somewhat far from that.
There are hundreds of different jobs in a company and a large part of them are repetitive, rule-defined tasks that do not build judgement. Invoice matching, data entry, routine reporting, format correction. Accelerate all the way there; in my view the erosion discussion contributes nothing in that space. What needs protecting is a small set; maybe one job in every ten. How a strategic decision is constructed, how the root cause of a problem is found, which assumption an analysis rests on. Every recommendation from here on applies in that area.
I think most companies' real problem today is never having made this distinction. Without the distinction you swing to one of two extremes: either you are cautious everywhere and lose on speed, or you ask questions nowhere and stay exposed to erosion.
Making the distinction is the most cost-effective move in this article.
What can we do as individuals?
Write your own answer first. Before asking the AI model, take two minutes to note your own hypothesis, then compare it with the output and explain to yourself where the difference comes from. A study conducted with radiology specialists in Frankfurt showed that physicians working with AI support went through a thinking process while comparing their own diagnoses with the system's output, and that this strengthened their professional accumulation.
Ask for a hint instead of an answer. This is the individual counterpart of the high school experiment. Instead of asking the model for a ready solution, ask for the next step, build your own path, and ask again when you get stuck. Same tool, two different outcomes.
Use the tool like a critic. Most of us ask the model for a better solution. One of the techniques BCG compiled suggests the opposite: ask the model to refute the solution. Questions such as "In which situation does this approach fail?", "Which is the weakest assumption in this analysis?", "Argue that this idea will not work" reveal breaking points that never appear in a plain prompt. The idea here: "Position the AI tool not as an expert who gives you answers, but as an opponent who challenges you". You still produce the answer, but you produce it more soundly.
Test your ownership. In an experiment in which MIT Media Lab monitored participants with EEG, the group using a large language model (LLM) fell behind the group not using the tool at neurological, linguistic and scoring levels over four months. According to the study's conclusion, those in this group struggled to quote from text they had written themselves a few minutes earlier. You can put yourself through such a test. Can you close a text you have delivered and explain its main argument in your own words? If you cannot, you have not yet mastered that work.
Keep a block where you do not use AI. Once a week, do a piece of work requiring synthesis from beginning to end yourself. If there is no place where the muscle works, you cannot know whether the muscle still exists either.
What should companies watch out for?
Individual discipline alone is not enough, because if the organization encourages the opposite, discipline will not be lasting and long-lived. In BCG's research only one in every ten companies has a strategy on this. There will be seven headings HR needs to bring into focus on this subject.
Competency architecture. Competency sets need to be redesigned according to erosion risk. The same at the behavioural indicator level: the phrase "analyses complex problems" today places both the person who genuinely analyses and the person who copies the output on the same line. An observable alternative should describe the behaviour of questioning the output's assumptions and, where needed, producing one's own framework. Without waiting for the next update, you can start by marking which item in the existing dictionary is exposed.
Mentoring. According to Gartner's analysis published on AI lock-in, by 2028, 40% of employees moving into a new role will be trained and coached by AI; today this rate is below 5%. The counterpart of the fast onboarding promise is fewer opportunities to learn from an experienced colleague. The structure Salesforce adapted from the pair programming idea in software offers a simple antidote: pairing someone comfortable with the tools and someone hesitant on the same piece of work. There is also an indicator for those who want to measure; how many hours a month does a junior employee spend on the same work with a senior?
Performance management. At CNIL, France's data protection authority, managers evaluate employees on their ability to challenge AI output; being able to use the tool is not an indicator on its own. It looks like a small change, yet it is a practice that will build culture in the organization: here, questioning is part of the job. The system also has a differentiation problem. Because AI brings an average employee's output closer to the level of a good employee, everyone's performance can start to look alike and the real capability gap does not show on the evaluation form. That is why measurement needs to place how the work was produced next to the work delivered.
Talent management and workforce norms. In BCG's study 53% of executives say the development of junior talent has slowed. The first rung of the career ladder always consisted of the same tasks: research, drafting, data cleaning, breaking the problem into pieces. Those were exactly the repetitions that build judgement, and this rung has now been handed to AI. When you remove a role, it is not only a cost item that goes; the ground on which the next role it feeds is grown goes too. The additional question to ask in workforce norm studies: if we remove this role, who will back up the role above it – who will grow into it?
Learning and development. While certification rates rise, no evidence appears that capability is rising. The approach the Aalto team recommends is to build automation-free learning environments, run regular workshops where employees solve complex cases together, and put in place periodic reviews that force people to justify the output. An Indian bank applies this with a tool-free session run across all functions on the first Friday of every month. Some companies place deliberate hallucinations and trap questions into training modules; an employee who encounters an error under controlled conditions directly acquires the reflex of questioning the output.
Recruitment and selection. If erosion is a risk, what we measure at the entry point must change too. How well a candidate masters the tool has ceased to be an easily differentiating criterion. What differentiates is their ability to evaluate an output placed in front of them and to frame the problem on their own. It is possible to build case interviews accordingly; giving the candidate a ready AI analysis and asking them to find the faulty assumption inside it says more than a classic problem-solving question.
Job design. This heading usually enters HR's agenda late, and to my mind it is the most decisive one. Once the decision of which work goes to the tool and which stays with the human is made, everything else is built on top of it. The rule here is to keep the sequences of work that build judgement with the human. Because AI agents can now carry out multi-step tasks from beginning to end, the loops that build skill can be squeezed out and disappear without notice. When designing the workflow, a second criterion needs to sit next to efficiency: at which point does this flow force the human to think?
Knowledge management. A large part of corporate memory was never written down; it was transferred from senior to junior while the work was being done. A third of the executives BCG interviewed reported a decline in mentoring and knowledge sharing; instead of consulting a colleague, employees go straight to the AI model. As this channel narrows, company-specific knowledge will start to erode too. The model tells you the sector's general truth; it does not tell you what is done differently in your company and why.
As we approach the end.
If I return to where I started, I think the task of embedding AI into the organization sits with HR. After this article I would like to add one more sentence to that: that task does not only contain rolling it out. It also contains knowing what is eroding, seeing in advance where it will start to erode, and taking precautions before it erodes.
This statement in BCG's report, I think, also summarizes the article;
"Human skills are not a fixed asset. They are a renewable resource that atrophies when unused and grows with deliberate practice."
The example that best explains this, I think, is a foreign language. You learn a language, you speak it fluently, you think in it. Then you don't use it for a few years. You do not forget the language structurally, you remember the words and you remember the grammar too, but you realize that the old fluency is gone and that you cannot express yourself as you wish; you hesitate while forming sentences, the word you are looking for comes late, you get tired while speaking. No one tells you "you lost your language", and you do not notice it in a single day. You understand it on the day you have to speak again.
Judgement is like this too. When unused it does not disappear, it starts to atrophy. And you only notice that atrophy on the day you need it.
Where we need to be careful here is being able to see the depletion while it is depleting. This does not require a complex system; there are three things you can look at today without spending any budget.
Doability without the tool. Pick your five most critical processes and ask each of their managers a single question: how many people do we have who can do this work from beginning to end without AI? If the answer is below two people, there is a fragility there. If you repeat the same question every year, you have a trend.
Junior exposure. How many hours a month does a junior employee spend on the same work with a senior? If this number is falling, the channel through which judgement is transferred is weakening.
One question in calibration. Add this to your next calibration session: which judgement and control point did this person build themselves while producing this result? The rate at which managers can answer this question tells you what the system sees today.
None of the three counts as a perfect measurement. But they are better than not measuring at all, and all three can be put in place very quickly.
If you are looking for a more holistic start, IC Consulting's HR Digital & AI Check-Up can give you a report on whether your existing competency, performance and development systems are able to see this risk, and you can schedule an evaluation meeting with me.