Artificial intelligence is making experienced professionals faster. It is also beginning to remove the work that made them experienced in the first place.
Across software, law, medicine, finance and other skilled professions, AI is taking over work once assigned to beginners: early drafts, analysis, diagnosis, financial models and routine coding. Much of that work was slow, and some of it tedious. Removing it can look like obvious progress.
Yet this was also where people learned their profession.
A junior did the work, made mistakes and had them corrected by someone who knew more. Then came another piece of work, and another. Over time, something changed in the person doing it. Experience accumulated. Judgment developed. The beginner learned to recognise when an answer that looked perfectly reasonable was wrong.
Most of today's senior professionals learned their craft before generative AI became part of the job. Their experience can conceal the weakness for years because the organisation continues to deliver and the numbers may even improve.
The problem comes later, when those seniors begin to leave and the people behind them are expected to take their place.
If a young lawyer has spent years checking machine-produced drafts rather than learning to build an argument, or a junior engineer has reviewed code without struggling through enough of it himself, time served is no guarantee that competence has been acquired. The same question now reaches medicine, finance and every profession in which judgment is built through repeated work under correction.
The Severance is not a case against artificial intelligence.
The gains are real, and there is no case for keeping work that no longer serves a purpose. The difficulty is identifying the work a junior still needs to do for himself, because without enough of it he may never acquire the judgment the job demands.
Constantine Leo Serafim has spent fifty years watching successive technologies change the workplace. This wave troubles him for a different reason. Earlier technologies removed labour and changed jobs. Generative AI can also remove part of the route by which a beginner becomes someone trusted to carry responsibility.
That leaves institutions facing a decision they have rarely had to make openly: which parts of a junior's work still need to be done by the junior, even when the machine is quicker?
The answer has to come from inside the profession. Experienced practitioners know which early tasks taught them to notice what others missed, where mistakes exposed weak thinking, and when correction began to change how they approached the work. Some of those tasks looked routine at the time. Their value only became clear later, when the habits formed there were being relied upon.
Much of it can now go. The mistake would be to assume that nothing is lost simply because the finished work still looks good. A junior who moved too early into checking machine-produced answers may remain useful and productive. What he may not be getting is enough practice finding his own way through a problem before someone, or something, gives him the answer.
The institution has to decide where a junior must still do the work unaided and who is responsible for correcting it. That choice costs time and may reduce an immediate saving, but it is part of the cost of producing people who can eventually carry the work themselves. Remove too much of that experience, and the organisation may find that the people coming through are not ready for the responsibility waiting for them.
The greater danger may take years to become visible. By then, AI may be better at the work than it is today, while too few people have acquired the experience needed to know when its answer is not enough.
Software can be bought quickly. A trained person cannot.