The AI boom is changing workplaces in myriad ways that extend well beyond efficiency gains. As companies automate more knowledge work, researchers warn of a worrying threat: the erosion of human skills.
This “de-skilling”, once seen as the natural shedding of obsolete tasks, can leave employees unable to perform essential functions when automation fails.
Few studies capture this as clearly as The Vicious Circles of Skill Erosion, published in 2023. That paper examined an accounting firm where reliance on automation fostered complacency, and eroded staff awareness, competence, and the ability to assess outputs.
When the system was removed, the firm realized its employees could no longer perform core accounting tasks.
The paper’s findings are more pertinent than ever in an era where AI tools are becoming ubiquitous.
I spoke with two of the paper’s original authors, Esko Penttinen (Associate Professor at Aalto University), and Joona Ruissalo (Post-doctoral researcher also at Aalto University), about the risks of skill erosion, why the issue is more urgent in the age of AI, and what businesses can do to prevent it.
- Your research talks about AI’s erosion of workspace skills. What motivated you to explore that subject?
Esko Penttinen (EP): In our research team, we pursue something that we call “phenomenon based” or “problem-based” approach to research. By that, we mean that we always start our research project with a practical problem that we encounter in “real-life”.
In this case, it was a serendipitous interaction with an informant in an accounting company who, in a side sentence, told us that an automation system had been removed from their IT architecture, revealing that their accountants’ skills related to the underlying business process had been eroded.
We found this observation made by the accountant fascinating and embarked on a case study to understand how this skill erosion manifested in the organization and how the erosion had happened. This truly was a revelatory case in the sense that it is extremely difficult to get access to this kind of failure case.
Organizations are typically reluctant to share their experiences regarding failures. We were very lucky and remain grateful that the organization let us study this phenomenon.
- What does it matter for every business and beyond?
EP: Our main finding points towards the delicate balancing act of handing tasks to technology while being mindfully engaged in the business process. This is what we depict in the figure in our paper (Figure 3 on page 1391).
We claim that most companies need to take their stance on this mindful conduction vs. automation reliance conundrum. These loops are by no means mutually exclusive, but we claim that it is very easy to go for the extreme; either fully automate something or then fully manually conduct the process.
The sweet spot is located somewhere in between these extremes. But this is easier said than done, as is shown in our case.
- Some might say that this is just evolution and that things have to change. How is this different?
EP: This is a good question and we struggle with this. For decades (and centuries for that matter), the objective of technological development has been to free human effort for more productive work.
There is this pro-automation argument, claiming that we should automate everything that can be automated so that human effort can be targeted to “higher thoughts”.
This applies in many arenas, accountants should not be manually entering invoice data into systems, but rather analyze the invoice data to provide insights to managers.
The flipside of this is that higher thoughts cannot be had without engaging in details.
By engaging in tedious manual work, employees often immerse themselves in nitty gritty details about the business process. And by this immersion, better insights can be gained.
- Given the quasi-evangelistic nature of the current AI ecosystem, can anything be done to mitigate this erosion?
EP: There are measures that organizations can take, for sure. We are writing a paper on this topic, hopefully getting it out next year.
In that paper, we stress the importance of technical and organizational control points. Instances of checks enforced to employees to check that they are in the “loop”, in other words, that they understand the actions taken by the automated agents or AI-tools.
Joona Ruissalo (JR): Other options are to organize recurrent workshops where employees work together to solve complex or unusual cases or even build automation-free training environments to raise awareness of potential gaps in skills and domain knowledge.
Also, to not “fall asleep behind the wheel”, organizations can run periodical audits in which they are asked for justification of outputs or implementing a nudging feature that regularly asks the human employee to validate an output with justifications.
In addition, adding explanation features to the essential systems they are using in their daily work allows to not just learn or quickly recap how the system operates to produce a specific output.
These measures should ideally be implemented in tandem to challenge employees to engage in reflection and critically evaluating AI outputs.
Actively maintaining organizations’ skill and knowledge capital puts them in a position where they can quickly adjust to external shocks where core systems can be taken down for an unspecified time and changing technological conditions to better co-evolve with new technological capabilities.
- One of the themes we discussed via email was prompt engineering (or composition as I put it). Can you draw a link between prompt engineering and the issue of skill erosion?
JR: Getting the prompt “right” is one thing, but it is quite another to evaluate those outputs.
These require different skill sets, but both necessitate competence in the problem domain, such as in accounting or software engineering, and lengthy exposure to the contextual intricacies to become efficient in composing the prompt and then validating the output you receive.
Of course, you can take ready-made prompts and let them spit out a response without scrutinizing them in depth, but where is the critical evaluation of the outputs and active reflection on why and how you are going about the process?
This is where the dynamics of skill erosion come to play: the issue of relying on the pre-validated prompts made by you in the past or someone else and repeatedly relying on those ceases your active engagement with the task where you no longer apply your skills and knowledge to the full.
As the prompts continue to produce the desired outputs, such as accurate financial information or lines of code, we run the risk of automation complacency where we become even more reliant on the generative AI’s outputs and are thrown to the fringes of being in the “loop”.
And as more time passes by, this is the moment where the issue of skill erosion might blow to our face: the prompt that produced the accurate output for a long time does not do so any longer as the underlying GenAI tool’s model changes (such as OpenAI forcing the move from GPT4 models to a single GPT5 model) or the software that builds upon the lines of code created with GenAI tools breaks.
If we have become complacent about conducting work mindfully and digging into the details, it is likely that skills have eroded over time.
Therefore, on an individual level, critical thinking and maintaining active reflection is essential as GenAI tools’ responses can at first look convincing, but as we know, looks can be deceiving as the responses can be suboptimal or partly hallucinated.
This issue is even more profound to junior employees who will likely have less chances to immerse themselves with the work context and facing less challenges to solve if they are mostly evaluating GenAI tools’ outputs – in other words, there are less chances to learn on the job to become experienced.
- Given the urgency and the clear risks associated with the phenomenon of skill erosion, why isn’t this issue pushed atop agendas?
EP: What makes this phenomenon tricky is its latent nature. If a company fully automates a business process, there are no problems as long as the system works.
This was true in our case organization as well.
The system was in a way “too perfect”, effectively optimizing the client companies’ optimization of their fixed assets. So why push something atop agendas if there are no problems?
Problems arise then when something goes wrong. In this case, it was a top management decision to discontinue the automation system.
The environment changed, leading to the discovery of the detrimental latent effects on employee skills. In some other contexts, it might be some other form of trigger that unearths the long-term impactful problems related to automation reliance.
- Anything else you want to discuss that wasn’t covered in the questions above?
EP: Which skills are such that should be retained and which skills can be forgotten or eroded? Drawing this line seems to be problematic.
Partly due to the changing environments. Something considered redundant now might not be considered redundant in the future.
We encourage companies to engage in scenario analysis, what are the possible and foreseeable alternative scenarios on organizational, technological, and environmental fronts?
How likely is it that an automation technology or an AI-tool that an organization has deployed suddenly becomes unavailable?
How likely is it that an environmental change impacts the required necessary skills in the business process that I am personally responsible?
What if our organization makes a strategic decision that impacts our IT infrastructure in a way that jeopardizes our IT?
These are the questions that we would like companies to consider.
