Every major technological wave arrives with the same apocalyptic scenario: this time, human labour will become truly obsolete and unnecessary. Since 2022, Generative AI has revived this old anxiety with a force I have not seen throughout my own career. It isn’t just that the technology arrived suddenly; it’s that its outputs(the text, images, and code it produces) feel unnervingly human.
It is tempting to look at the headline numbers, like the percentages of jobs at risk, and treat the debate as black and white issues of survival or destruction. But as I have found in some of my recent research, this framing results in companies and individuals becoming paralysed with panic, when what they really need is preparation and pro–activity.
We should re-frame the debate around AI and the workplace by asking not, “how many jobs will disappear?” but rather: which tasks can be shared and reallocated between humans and machines, at what speed, and what will the consequences be for skills development, pay, and job quality?
In labour economics, the most misleading way to think about AI is to treat a job as a single, indivisible block. This mental model is exactly what produces the end of work hysteria. In reality, an occupation is a made up of multiple, varied tasks. While AI may be highly efficient at substituting some of these specific tasks, it is rarely a one-for-one replacement for an entire occupation. In fact, only about 20% of occupations have most, not all, tasks exposed to AI.
Why tasks, not jobs, are the real unit of change
This difference between tasks and occupations is the key to understanding our future. Historically, very few job categories have actually disappeared. Instead, they have undergone an internal transformation. The real disruption happens inside the job and the parts being automated, and what resurfaces as a scarce human advantage.
In my recent studies of the French labour market, I used task-based modelling to score the substitution potential of various roles. Interestingly, the pattern of this wave differed from the automation of the past. It is not just routine manual work that is “exposed” this time but high-skilled, cognitive roles which require analysis, coding, reporting, and design.
But there is a further caveat: exposure is not the same as impact. A task can be technically automatable and yet still not be automated due to economic feasibility and organisational friction. Implementing AI is expensive and risky. AI errors are famously unpredictable, and they can be confidently wrong in ways a human would not be. This creates a cost of failure that many companies are currently underestimating.
This brings us to what I call the “last mile” of production. AI can be remarkably impressive until it reaches the point where an output must be context-aware, compliant, and aligned with organisational reality. At this stage, human expertise becomes more critical than ever. Without it, the time gained through machine use risks being lost later in the process of verifying and correcting.
The hidden limits of automation in practice
Transitioning into a new era of work creates a massive responsibility for HR and managers and business leaders.
We are entering a time where machines are mirroring human skills: the capacity to assess quality, detect errors, and ask the right questions. However, if some of these tasks, especially those that we consider entry-level tasks, are automated, how will junior employees learn?
For decades, careers have begun with routine tasks. These weren’t glamorous, but they were the training grounds where juniors absorbed knowledge and built the judgment required for senior roles. If we remove these training responsibilities, we may accidentally destroy the pipeline of future experts.
We risk a scenario where junior workers become mere supervisors of machine outputs without ever developing an underlying understanding of their profession. This isn’t just an HR problem; it’s a long-term threat to organisational competence.
There is, however, a more optimistic path. Data from a recent study I conducted with Philippe Aghion at HEC Paris, shows that firms adopting AI between 2018 and 2020 saw employment and sales increase by five percent. Much like the spread of ATMs in banking, the automation of specific tasks allowed for an expansion of overall activity. The productivity gains helped firms grow, shifting labour toward functions that the machines simply couldn’t do.
Protecting skills while embracing productivity gains
The path forward for business is not to wait for the storm to pass, but to actively manage the reallocation of skills. HR leaders must start by mapping tasks rather than jobs and explicitly decide which will remain training tasks for junior employees, even if a machine could do them faster.
We must treat job quality as a key performance indicator; if AI only leaves humans with tasks to edit and check, work will become cognitively demanding but intrinsically hollow.
The panic around AI taking jobs is misplaced but leaders and HR managers do need to pay attention. Although we are not witnessing the collapse of employment, we are experiencing a fundamental reallocation of human effort and focus.
We must make sure that as we hand routines over to the machines, we also liberate individuals’ time and energy for the tasks that require what no algorithm can provide: responsibility, empathy, and genuine judgment.
A specialist in economic growth and innovation, Antonin's research focuses on the long-run determinants of GDP evolution, whether through studying sources of heterogeneity between firms, the creation and diffusion of innovations, or the dynamics of productivity.
Antonin received his Ph.D. from the Paris School of Economics in 2018 and is a graduate of Ecole Polytechnique. Before working at HEC, he was a research economist at the Banque de France.
