Without apprentices, who will know how to judge?

Picture a small accounting firm in Sion. A file arrives with its invoices, its statements and an entry that does not balance. The apprentice searches. He matches two amounts, gets the period wrong, starts again. His trainer sits down beside him. She shows him what he should have been looking at, then asks him why he did not see it. The file moves forward slowly. Someone is learning a trade.

In another version of that scene, a tool has already matched the documents and proposed the correction. The trainer checks, approves and moves on to the next client. The file is handled faster. To find out what the apprentice has learned, someone would now have to ask him a question.

This scene extends a difficulty left open in my essay. In it I defend the value of experienced practitioners, able to frame a request and to challenge a generated answer. But that experience had to be formed somewhere. If we transform the work that produced it, we also have to rebuild the occasions for acquiring it.

The commercial apprentice, the young accountant and the legal trainee follow different paths. All of them, however, have to move from an instruction to an understanding. Why is this document missing? Why does this clause deserve a reservation? At what point should you stop searching alone and ask for help? Judgement is also built in those hesitations, when a competent person helps to resolve them.

Routine tasks sometimes had two functions: moving production forward and making that progression possible. Their pedagogical value was not always visible in billed time. It appeared later, when the beginner recognised an anomaly without anyone pointing it out.

Not all tedious work deserves to be preserved, of course. Copying the same information a hundred times may teach almost nothing. Examining a few well-chosen files, understanding their differences and explaining a mistake can teach a great deal. Automation forces us to make that sorting more consciously. Every task removed should be matched by a precise question: which skill did it allow someone to acquire, and where will that skill now be learned?

AI can itself contribute to this learning. A study by Brynjolfsson, Li and Raymond, published in 2025, follows the introduction of an assistant among 5,172 customer support agents. Productivity rises by 15% on average, with larger gains among the least experienced. The authors also observe signs of learning that persist when the tool becomes temporarily unavailable. [1]

These results concern one company and one specific activity. They do not directly describe learning in a Valais accounting firm. They do, however, forbid treating AI as necessarily unfavourable to beginners. Depending on how it is built into the work, it can give them earlier access to practices they would have taken a long time to encounter.

A distinction then becomes essential: obtaining a result, and understanding what makes it acceptable. A beginner may receive an excellent proposal without knowing under what circumstances it would stop being right. Training has to let him recognise its limits, look for a source and explain his choice. That capacity requires opportunities to practise, with support that changes as he progresses.

The problem becomes economic when we ask who pays for that support. During training, an apprentice can already produce real value: preparing documents, handling operations suited to his level, freeing up time for colleagues. The net cost of his presence depends on that contribution, set against his pay, the social charges, the resources committed and the supervision time. This calculation has to cover the whole course of training: what he brings changes as he becomes autonomous.

If AI absorbs his useful tasks while supervision remains necessary, that balance deteriorates for the company. Asking it to maintain the same effort in the name of the next generation does not solve its problem. The cost it bears, the design of the position and the progression towards a productive contribution all need to be reviewed. Pay is one component of that equation; the trainer's time and the learning situations to be rebuilt must also be costed. And where AI increases the value produced by the beginner, the diagnosis may be quite different.

To that cost is added the uncertainty of the return. A person may join another company once trained. The firm that paid for the support will then not necessarily benefit from her work once it has become expert. A viable model has to take that mobility into account, without assuming that gratitude will secure loyalty. If everyone prefers to recruit already trained professionals, everyone is counting on an investment made elsewhere, which no one will have sufficient interest in bearing.

This risk still has to be measured; it does not allow anyone to announce the disappearance of apprenticeship places. But it is enough to raise a collective question. A territorial strategy resting on the skills available today has to look at how they will be renewed. Attracting experienced workers can help Valais. It does not remove the need to train those who will take over here.

The canton already has training companies, vocational schools and industry associations. These actors could begin by examining together what automation changes in a few training paths. Which tasks are leaving beginners' hands? Which skills develop faster? Where does a trainer have to intervene more? A modest observation, carried out in several volunteer companies, would be more useful than a general doctrine settled too early.

In the firm we have imagined, certain files could be set aside for an explicit progression. On a case suited to his level, the apprentice would begin by formulating his own hypothesis. He would then compare it with the tool's proposal. The trainer would discuss with him the gaps, the checks required and the final decision. The work done would then have a double value, productive and formative, chosen from the outset.

AI could prepare variants of an exercise or offer additional explanations, under the trainer's control. Rare cases, ambiguous documents and plausible errors would become material for learning. It would also be possible to check regularly what the young person can explain and resolve on his own, in order to distinguish the progress of his competence from that of the system he is using.

What remains is to make this effort financeable. Some of the hours saved could cover supervision, provided the gain is real and available. Several small firms could share anonymised cases and training sessions. Sector-level funding could also spread part of the cost between companies that train and companies that recruit the people they have trained. If one phase becomes essentially educational, the place of public funding deserves to be examined as well. These options require costing: net cost per training path, growth of the productive contribution and mobility after training.

I have written that experience was becoming a strategic asset as AI spreads. Its transmission therefore has to find an explicit economic model. A company must be able to train without betting its whole investment on one person's future loyalty. A young person must be able to acquire a useful skill and make a living from it. Holding those two requirements together becomes a matter for the sector and for the territory.

Let us return to our file, on a desk in Sion. A few years later, the trainer will retire. The tool will probably have changed. There will still be ambiguous documents, anxious clients and decisions to explain. The question will then be very concrete: who will she be able to hand the file to?

[1] Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond, Generative AI at Work, The Quarterly Journal of Economics, vol. 140, no. 2, 2025, pp. 889–942.

See also · Chapter 10 — Trust, Legal, Consulting Firms

See also · Chapter 12 — Training, Reskilling, Alpine Campus