Qualified cognitive tasks, long considered beyond the reach of automation, are now on the front lines. Training budgets are contracting at the precise moment when the transition would require them to expand. The 2027 trade-off is clear: tax the productivity gains from AI to fund retraining, or leave workers still in post bearing the cost of automating their colleagues’ work.
The frontier of automation has shifted upward
With AI, it is complex, qualified cognitive tasks that are now exposed. A joint study by Coface and the Observatory of Threatened and Emerging Jobs (OEM), published in March 2026, documents this shift [2]. This reverses the logic of previous waves of automation, which struck intermediate and repetitive jobs.
In 2026, 3.8% of French jobs are vulnerable to generative AI [2]. Over a two to five-year horizon, this rate rises to 16.3%, or nearly 5 million positions [2]. The study specifies that this indicator “reflects task exposure, not their destruction.” Among the 923 professions analyzed, 120 present a high level of exposure, with more than 30% of their tasks potentially automatable [2].
The distribution by income level is counterintuitive. High salaries face the strongest exposure: the top 10% earners show 22.1% exposure in 2026 [2]. The most qualified jobs—engineering, IT, law, creative professions—are among the most exposed. Generative AI excels precisely in analysis, synthesis, writing, and documentary research, which form the core of these professions.
Overall unemployment holds, but the first rungs disappear
Labor markets showed unexpected resilience in 2025 and early 2026. In May 2026, the average unemployment rate in the OECD zone stood at 4.9%, close to its historic lows [3]. The employment rate reached a record 72.1% [3].
These aggregates mask a transformation already operating at a finer level. The effects of AI are beginning to appear in early-career jobs in the most exposed professions [3]. In January 2026, the IMF confirmed that automation strikes junior positions two to three times more than managerial posts [8]. Tasks once entrusted to entry-level profiles—simple content writing, documentary analysis, first-level data processing—are the first affected.
French data confirm this pressure on the youngest. In 2026, 78% of young professionals found employment within six months of finishing their studies [3]. It was more than 85% after the health crisis [3]. Overall employment holds steady, but the first rungs of the professional ladder are becoming scarce. Axelle Arquié, cofounder of the OEM, raises an additional dimension: if young graduates anticipate the devaluation of their qualifications, the investment in higher education could lose its individual economic rationale [1].
Agentic AI, a rupture of another kind
The jump from 3.8% to 16.3% exposure between 2026 and the two to five-year horizon is explained by a qualitative change in the nature of AI being deployed [2]. So-called agentic AI replaces an entire workflow, whereas AI assistance replaces only an isolated tool. This distinction alone explains the scale of the projection.
AI is a general-purpose technology in the economic sense: it touches all sectors and modifies all production processes. Carl Benedikt Frey argues, in How Progress Ends published in 2025, that AI will have the same effect on qualified workers that the deindustrialization of the 1970s had on blue-collar workers [9]. His decisive contribution is to show that this outcome depends on the robustness of the welfare state in absorbing job displacement.
The bulk of AI adoption in companies is currently oriented toward automation: doing the same tasks faster, cheaper, with fewer people. Acemoglu, Autor, and Johnson formalize this finding in a February 2026 working paper [4]. Their framework distinguishes five types of technological change. Only the creation of new tasks generates demand for new human expertise. AI is currently deploying mostly in the other categories [4].
France starts with an adoption handicap that complicates diagnosis. The ECB’s SAFE survey in the fourth quarter of 2025, conducted among 625 French companies, reveals that 23% reported moderate to significant AI use [5]. The average for the euro zone was 39%. Germany reached 46%, Spain 44% [6].
INSEE published in July 2026 that 18% of French companies with ten or more employees used at least one AI technology in 2025 [7]. This was 10% in 2024, and 6% in 2023: the rate tripled in two years [7]. Companies using AI now account for 66% of revenue and 59% of employment in the field studied [7]. The lag of small and medium-sized enterprises creates a bifurcation between two labor markets: that of companies adopting AI, and that of the others.
Training budgets contract at the worst moment
The 2026 budget for France Compétences provides for 12.078 billion euros in spending. That is nearly 1.5 billion less than in 2025 [10]. Apprenticeship absorbs the bulk of the cut: its envelope falls from 9.3 to 8.2 billion euros [10]. The budget for the personal training account, the training rights accumulated by each employee, drops by 650 million euros, to 1.31 billion [10].
The average budget per beneficiary falls from 3,250 to 2,870 euros in 2026 [11]. The average length of training falls from 3.8 to 3.1 months in the same year [11]. On the ground, there is a return to short formats with little qualification, at the very moment when professional transitions require programs of 12 to 18 months.
The fiscal shock is the most underestimated dimension. Rapid automation of qualified functions first strikes jobs that contribute most to tax and social revenue bases. In 2026, the OEM estimates that the most exposed positions weigh heavily in the formation of government revenues [1]. A rise in unemployment benefits combined with a decline in contributions creates a double shock for public finances. Without an organized fiscal transition, the balances in social protection financing are destabilized.
What the AI Act, which came into force in August 2026, does not settle
The European AI Act has governed high-risk uses of AI since August 2, 2026 [12]. When AI affects recruitment, credit, health, or performance evaluation, companies must document, trace, and guarantee the transparency of their system. This is a real achievement.
The regulatory framework addresses the individual risk of each system. It does not govern the macroeconomic transformation of employment structure [14]. Robust rules on the uses of AI systems do not cover the labor market effects of a general-purpose technology, as Jean Tirole underscores in his work on the regulation of digital markets [14].
One dimension fades from French debate. Models, data, graphics chips, and cloud infrastructure are the property of large American groups. In 2024, more than 70% of European companies’ cloud spending went to extra-European providers [15]. French employment’s exposure to AI is therefore an exposure to tools whose development orientations are decided beyond the reach of European policies, as shown by Stéphane Grumbach’s work on the geopolitics of data [15].
Direct deployment, extend training, tax the gains
Three levers flow from this causal chain, and they are not equivalent.
The first is to direct the deployment of AI. Automation is one choice among others that is technologically available, as Acemoglu, Autor, and Johnson show [4]. An AI that augments human capacities creates value without destroying jobs.
It helps a nurse interpret complex data. It allows a lawyer to handle more cases without eliminating colleagues. Public procurement, innovation aid criteria, and access rules for public data can tilt deployment orientations. This is a lever distinct from simple risk management under the AI Act.
The second is to redirect training budgets toward long-term skills. The contraction of 1.5 billion euros at France Compétences in 2026 is a sequencing error [10]. Bringing the average duration of training down to 3.1 months [11], when transition requires retraining of 12 to 18 months, allocates spending where it is least effective. Long-term qualified training is the missing lever of the French system, which Bertrand Martinot documents in his work on labor policy [13]. The budget constraint is real; it demands choices in spending, not a uniform reduction.
The third lever is the most structuring for 2027: create a contribution on AI productivity gains to finance transitions. Creative destruction does not spontaneously redistribute its gains [14]. The IMF insists on the possible complementarity between AI and labor, provided that workers are trained, work is reorganized, and a share of benefits is redistributed [8].
The best-paid positions contribute most to the general social contribution, income tax, and employer contributions. These are also the first exposed, as Arquié documents [1]. If these positions disappear or shrink, the financing of social protection depends directly on how AI productivity gains are taxed. Making workers still in post bear this cost through payroll contributions amounts to having them finance the automation of their colleagues’ work. A contribution on profits drawn from automation, distinct from capital taxation in the broad sense, would allow an increase in a dedicated transition fund.
This contribution is the transitional instrument of a broader shift. The financing of social protection must expand beyond wage contributions alone. This agenda is detailed in the post “Seniors locked out of employment, shrinking social revenue base.”
Sources
[1] Axelle Arquié, “The double shock of AI: employment and taxation,” L’Économie politique, no. 110, 2026/2, Alternatives économiques, https://shs.cairn.info/revue-l-economie-politique-2026-2 (accessed 25/08/2026).
[2] Coface and Observatory of Threatened and Emerging Jobs (OEM), “Jobs, skills, value: what AI is disrupting,” March 2026, https://www.coface.fr/actualites-economie-conseils/emplois-competences-valeur-ce-que-l-ia-est-en-train-de-bouleverser (accessed 25/08/2026).
[3] OECD, Employment Outlook 2026, OECD, Paris, July 2026, https://www.oecd.org (accessed 25/08/2026).
[4] Daron Acemoglu, David Autor, and Simon Johnson, “Building Pro-Worker Artificial Intelligence,” NBER Working Paper no. 34854, Hamilton Project, Brookings Institution, February 2026, https://doi.org/10.3386/w34854 (accessed 25/08/2026).
[5] European Central Bank, SAFE Survey (Survey on the access to finance of enterprises), 4th quarter 2025, Frankfurt, 2025, https://www.ecb.europa.eu (accessed 25/08/2026).
[6] Banque de France, “French Companies: Is there an AI adoption gap?”, research note, May 2026, https://www.banque-france.fr/en/publications-and-statistics/publications/ai-adoption-gap-among-french-firms (accessed 25/08/2026).
[7] INSEE, “Information and communication technologies in companies in 2025,” INSEE Première no. 2120, July 2026, https://www.insee.fr (accessed 25/08/2026).
[8] IMF, World Economic Outlook Update, January 2026, Washington D.C., https://www.imf.org (accessed 25/08/2026).
[9] Carl Benedikt Frey, How Progress Ends: Technology, Innovation, and the Fate of Nations, Princeton University Press, September 2025, https://press.princeton.edu/books/hardcover/9780691233079/how-progress-ends (accessed 25/08/2026).
[10] France Compétences, 2026 Budget Forecast, adopted November 27, 2025, https://www.francecompetences.fr (accessed 25/08/2026).
[11] DARES, 2026 Yellow Book: Vocational Training, annex to the 2026 Budget Bill, Ministry of Labor, February 2026, https://dares.travail-emploi.gouv.fr/donnees/le-jaune-budgetaire-sur-la-formation-professionnelle (accessed 25/08/2026).
[12] Regulation (EU) 2024/1689 (AI Act), Official Journal of the European Union, July 12, 2024; obligations of high-risk deployers, article 26, applicable August 2, 2026, https://eur-lex.europa.eu (accessed 25/08/2026).
[13] Bertrand Martinot, Work is the Solution, Les Belles Lettres, 2025, https://www.wikiberal.org/wiki/Bertrand_Martinot (accessed 25/08/2026).
[14] Philippe Aghion, “Resetting the Innovation Clock: Endogenous Growth through Technological Turnover,” hearing before the Economic Affairs Commission of the National Assembly, January 2026, https://www.assemblee-nationale.fr/dyn/17/comptes-rendus/cion-eco/l17cion-eco2526012_compte-rendu.pdf (accessed 25/08/2026).
[15] Stéphane Grumbach, “Digital growth, energy needs, environmental impacts and security concerns,” 2024, https://shs.cairn.info/publications-de-stephane-grumbach–97659 (accessed 25/08/2026).
