从安全到暴露：人工智能如何重新绘制工作地图

FOCUS By Axelle Arquié, Economistand Co-founder of theObservatoiredes emplois menacés et émergentsAurélien Duthoit,Senior SectorEconomist, based in ParisGaleran Subileau,Junior Economist The Next Automation Frontier:A Scenario Map of AI Labour Exposure EXECUTIVE SUMMARY Three years after the release of ChatGPT, AI seems to be everywhere in corporate discourse, and yet still almostnowhere in aggregate labour-market statistics. The first effects are however beginning to appear, mainly at themargins, among the younger, entry-level workers of the most exposed occupations and industries. This apparentparadox may simply reflect timing, with the current phase of AI adoption still largely focused on tools that assist workwithout reorganizing it, and the current evidence capturing early signals rather than the main shock itself. We believethe rise of agentic AI systems, that is, systems capable of planning, coordinating and executing multi-step workflows,could accelerate and amplify the disruptive impact of AI on labour markets by moving from assistance on isolatedtasks to the execution of entire workflows. To move beyond these early signals, the Observatoire des Emplois Menacés et Émergents (OEM) developed a task-basedframework to measure the technical exposure of occupations to successive phases of AI development. While manyexisting studies rely either on expert judgment or one-off assessments produced by large language models, the OEMapproach breaks complex tasks into elementary actions, applies explicit and reproducible scoring rules and projectsexposures along successive technological phases rather than a single point in time. The result is a granular, reproducibleand prospective way to assess how advances in AI could reshape occupations, industries and labour markets. Together with the OEM, Coface contributed to extending this framework by developing a weighting method fortasks, refining further both the prospective automation scenarios and the notation rules elaborated by the OEM, andby broadening the empirical scope of the analysis. This framework is deliberately gross and supply-side: it measurestechnical exposure to automation, not net job destruction. It abstracts from demand dynamics, from potential newtask creation and from frictions that may slow deployment. l l l Our results point to a different automation pattern from earlier technological waves such as industrial robotics of the1990s and 2000s and computer software. Rather than concentrating on routine, middle-skill work, the automationpotential of AI in its agentic form is highest in occupations made up largely of cognitive, non-routine tasks oftenperformed by more highly educated and better-paid workers. This is especially visible in engineering & computationaloccupations (29% of task content at risk), legal & financial, creative & content occupations (27%), and management& administrative roles (24%). At the national level, automation-exposed task content ranges from around 12% inTurkey to close to 20% in the UK, with countries such as France (16%), Germany (17%) and the United States (17%)lying closer to the middle of the distribution. Because the occupations most exposed are among the most central toincome formation, tax revenues and value creation in advanced economies, the consequences are likely to extendwell beyond employment alone. AI could create a significant redistribution of income away from labour and towardcapital. If employment losses in exposed occupations are not fully offset by new job creation, wage income wouldcome under pressure while productivity gains would accrue first to companies deploying AI. Part of these gains maythen be captured further upstream by the technology companies controlling the core infrastructure of the AI valuechain. This raises, for many countries with public finances heavily dependent on labour taxation, a double risk oferosion of the domestic wage tax base and partial leakage of capital income toward foreign countries where AI profitsare concentrated. The implications also extend beyond employment and include a rethinking of education systems as AI begins to testthe labour-market value of qualifications and diplomas and shift the premium toward judgment, adaptability and AI-complementary skills. They also include stronger supply-chain and geopolitical dependencies as AI becomes a criticalinput to production while its key assets (semiconductors, data centers, models) remain highly concentrated across asmall number of companies and countries, thereby increasing exposure to external shocks, regulatory frictions andstrategic vulnerabilities. While these results should be read with caution since the translation from the gross technical exposure of tasks to thenet employment effects is neither immediate nor mechanical, the underlying questions would remain even undera more gradual or limited disruption scenario. The framework should therefore be understood not as a forecast, butas a structured map o