The scene repeats itself in the businesses we work with. A task once handed to the newest arrival — reworking a document, checking data, producing a first draft — is now done by a tool, better and faster. Nobody has been made redundant. The business has simply stopped needing to hand that work to someone starting out. The “AI and jobs” debate misses that shift.
That is the real question of AI and jobs, and the 2026 edition of Stanford’s AI Index makes it possible, for the first time, to look at it with figures rather than opinions (Stanford HAI, AI Index 2026). They do not say AI destroys jobs. They say something more precise, and more awkward for a small business.
AI and jobs: the signal is not job destruction
The most quoted figure in this edition is American, and that should be said before quoting it. Using payroll data, researchers find that employment of software developers aged 22 to 25 has fallen by nearly 20% since 2024, while employment of more experienced developers has kept growing (AI Index 2026, p. 11). The body of the chapter gives a slightly more cautious measure, around 16% for that age group in the occupations most exposed to AI (AI Index 2026, p. 222).
Two cautions apply. The French labour market is not the American one: different contracts, different mobility, different hiring cycles. And correlation is not cause — the US technology sector went through other shocks over the same period. The report itself does not conclude causality. It notes a coincidence, and it is a troubling one.
Troubling because the decline is not randomly spread: it appears where AI-related productivity gains are best documented (AI Index 2026, p. 11). In other words, this is not a sector doing badly. It is a level of position disappearing in sectors that are doing well.
Nothing requires waiting for a French study to know where you stand. The figure is in your own records: how many people have you hired at the start of their career over the past three years, and how many over the three years before? If the curve is falling without any decision having caused it, someone has already decided for you, task by task.
What AI replaces is the learning exercise
Look at what a beginner actually does in their first two years, in any office job. They rework existing documents, check data, produce first drafts that someone else will correct, sit in meetings they only partly understand. That work has two functions: it renders a service to the business, and it manufactures a professional.
AI is precisely excellent at the first function, and has no effect at all on the second. The report measures, for instance, 26% more contributions from developers equipped with a coding assistant (AI Index 2026, p. 174). The gain is real. But if those contributions are the ones a junior would have written while learning, this year’s gain is paid for in missing skill five years from now.
The report names this risk without settling it: heavy reliance on AI may carry a learning penalty, that is, slow the development of skills over the long term (AI Index 2026, p. 174). This is not an established finding, and should be presented as such. It is a documented concern, raised by researchers who also measure very real gains.
The phenomenon is in no way specific to software, even if that is where it is best measured. In accountancy the first exercise was data entry and reconciliation; in law firms, research and the summary note; in agencies, the first draft of a text or a layout. Everywhere, learning travelled through work that could be handed over safely because it would be reviewed. That is precisely the category of work AI does well.
Your teams do not see what you see
One last figure explains a lot of difficult meetings. Asked about AI’s effect on how people do their jobs, 73% of experts expect a positive effect — against 23% of the general public (AI Index 2026, p. 12). Fifty points apart. A management team announcing an AI project speaks from the first group to people who belong to the second.
In a small business, that gap is not settled with a memo. It is settled by saying what will not change: who stays, which tasks remain human, and what is now expected of junior roles. As long as those three answers are missing, everyone fills them in with what they fear.
The conversation opens better with a question than with an announcement. Ask each person which part of their work they would happily hand to a tool, and which they refuse to let go. The answers are almost always finer than management imagines: teams distinguish very well between the painful task and the task that keeps them standing. It is that distinction, stated by the people doing the work, that should guide the choices.
What a manager can decide right now
Keep entry-level positions, first, and own that as an investment rather than an inefficiency. In a small business, every senior is hard to replace and hard to recruit: automating junior work with no transmission plan amounts to betting on the recruitment market five years from now.
Redefine the junior task, next. Producing the first draft teaches little now, since the tool produces it. Checking, criticising, arbitrating, going after what is missing: that is what can still be learned, and it is more demanding than before. A junior who methodically reviews a model’s output learns the trade faster than one who used to copy — provided someone reviews their review.
Organise transmission, finally, because it will no longer happen by itself. It used to travel through shared work; it will now travel through explicit gestures — one case followed end to end, one commented review each week, one error recounted to the team. The choice of tasks handed to AI then arises in the same terms as in our article on the limits of AI, with one extra criterion: does this task train someone?
We have set out elsewhere why AI accelerates a skill but does not replace it, and what its use costs without showing straight away. The Stanford report says nothing different, with data to back it — and the article that opens this series draws the overall consequence.
| What a beginner used to do | What the tool does with it | What trains them now |
|---|---|---|
| Produce the first draft | It produces it, faster and often better. | Judging that draft: what is missing, what is wrong, what must go. |
| Check and copy data | It extracts and copies without tiring. | Defining the checking rule, and handling the cases the tool flags as doubtful. |
| Look up a precedent, a standard, a file | It finds it in seconds, sometimes beside the point. | Knowing how a solid source is recognised, and finding the missing one. |
The question in the title is not rhetorical. If AI takes the tasks through which people learned, someone has to decide how people learn now — and that someone is neither the software vendor nor the labour market. It is a decision for the person in charge, it is taken early, and it costs little while there are still juniors to benefit from it.
Common questions
Which jobs are threatened by AI?
The report does not speak of threatened jobs but of automatable tasks, and observes a decline in entry-level employment where productivity gains are documented — in the United States, among developers aged 22 to 25. That is not a prediction, it is a correlation.
Will AI replace junior developers?
It already replaces part of their tasks, the ones through which they learned. The risk is not running short of juniors: it is running short of seniors in five years, for want of having organised learning differently.
How do you train your teams on AI?
On a real process rather than in a classroom. Everyone should be able to answer three questions: which tasks am I allowed to use it for, how do I recognise a doubtful answer, and who do I flag it to. The rest is learned by correcting real cases.
Source: Stanford HAI, Artificial Intelligence Index Report 2026, published on 29 June 2026 under a CC BY-ND 4.0 licence. Page numbers refer to the PDF file. The employment data cited is American, drawn from payroll records analysed by researchers; the report describes a correlation and does not conclude causality. This article is an analysis by Maeliom; it is neither a translation nor an adaptation of the report, and is neither affiliated with nor endorsed by Stanford University.