代理型AI：自主运行型经济？

16 September 2025 Allianz Research Content Page 3Executive Summary Page 4-6The rise of agentic AI Page 7-8 In tech we trust: Economic opportunity andsocial costs of agentic AI Page 9-14 From co-pilot to autopilot: Substitutabilityof labor for AI capital Page 15 New tools, new policies: policy levers thatmatter for demand Page 16Annex ExecutiveSummary •Agentic AI – capable of autonomous decision-making and executing complextasks without human intervention – marks a significant leap in artificialintelligence, with the potential to profoundly reshape economies andlabor markets across the globe.The economic promise is considerable, withprojected global gains of USD2.6-4.4trn over the next two to five years, drivenby productivity and innovation. But the key question is whether this productivitygrowth will come at the cost of widespread unemployment. Historically, the laborshare of national income remained relatively stable throughout most of thepostwar period, suggesting that technological advancements complementedrather than replaced labor. In recent years, however, this trend has shifteddownward, raising concerns about the increasing substitutability of labor with AIsystems. Compared to earlier AI technologies, including generative AI, agenticAI poses a greater risk to employment, with recent estimates suggesting that upto 60% of jobs in advanced economies and 40% of global employment could beeither augmented or automated by AI. The scale of this disruption could mirrorthe transformative effects of the Industrial Revolution, fundamentally challengingexisting models of labor, income distribution and economic growth. Patricia Pelayo-RomeroSenior Economist, Insurance & ESGpatricia.pelayo-romero@allianz.com Katharina Utermoehl,head of thematic and policy researchkatharina.utermoehl@allianz.com •To assess the relationship between labor and capital, we analyze labordemand across several countries.While in Germany, Spain, Italy, and Poland,gross fixed capital formation and labor appear to be clear complements, withPoland and Italy showing the strongest positive effect Austria, France, and theNetherlands show no statistically significant relationship. However, when focusingon assets that serve as AI proxies, such as software and R&D, a substitution effectemerges. In all countries studied, increased software investment is associated withreduced employment, ranging from 0.22% to 0.29% per 1% increase in investment.R&D investment also shows a negative impact on employment, though smaller(0.01% to 0.08%), and not statistically significant in Italy and Spain. •Software investments are found to reduce labor in all industries except foragriculture, with the strongest decrease in finance and real estate.Gross fixedcapital formation is positively correlated with labor demand across all industries,increasing employment by 0.15% to 0.35% per 1% investment. The strongesteffects are in agriculture, arts, construction and manufacturing. However, forsoftware investments we find that a 1% increase reduces labor in all industriesbar agriculture, from 0.04% to 0.18%. with the strongest decrease in financeand real estate. Similarly, an increase of 1% in R&D investments tends to reducelabor demand across most industries, with the greatest substitution in real estate(0.30%), finance (0.30%) and ITC (0.17%). Our findings indicate that AI, and inparticular agentic AI, may drive more substitution than previous technologicalwaves. •Against this backdrop, a holistic public policy approach will be essential tomitigate labor market disruptions, redistributing displaced workers into newoccupations and industries through retraining and reskilling, and incentivizingcompanies to hire displaced workers.The unique risks posed by the pace, scaleand nature of AI-driven job displacement also justify the creation of new formsof social initiatives to share AI gains with those affected, such as cash transfers,AI displacement insurance or a universal basic income, financed by an AI-usagecontribution system, a minimum corporate tax like that promoted by the OECD orpreemptively tackling the issue with employee profit-sharing programs.33 The rise of agentic AI Since the introduction of ChatGPT in November 2022, AIcapabilities have only accelerated.Figure 1 demonstrateshow artificial intelligence systems have equaled or evenrivaled the human benchmark. In fact, with GPT-5, OpenAIclaims that the new model for ChatGPT has been boostedto “PhD level”. The researchers behind AI 2027¹ and theCEOs of OpenAI, Google DeepMind and Anthropic haveall predicted that if the developmental trend continues,artificial general intelligence (AGI) will arrive within thenext five years. stick to a travel budget for a family holiday and it would,given the appropriate set-up, work as a travel agentand organize transportation tickets and hotel bookings,besides providing a suggested schedule. Likewise, it couldbe programmed to carry out operations in customerservice,