AI, human labor, and the task frontier: automation, complementarity, and the net effect of Artificial Intelligence on employment

The task-based analytic framework suggested by Acemoglu and Restrepo (2018, 2019) was developed to study machines that replace workers. This paper studies how the whole picture changes when the technology works with humans rather than replacing them. We add a complementarity zone to the framework − the part of tasks where AI predictions make human judgment more accurate rather than replacing it − in the spirit of the work by Agrawal, Gans, and Goldfarb (2019). This model emphasizes four main channels through which AI influences jobs: displacement (−), complementarity (+), productivity (+), and reinstatement through new tasks (+). As far as the data is concerned, we rely on an Eurostat dataset that gives the breakdown of AI adoption according to technology types across 27 EU member states (2021, 2025) as well as regions (59 NUTS 2 regions). We calculate the level at which countries are exposed to “automation-type” AI and “complementarity-type” AI. The results highlight three main points. Complementary AI (text mining, natural language generation, machine learning) is being adopted 2.6 times faster than automation AI (robots, RPA). Regions highly exposed to the latter are losing jobs (β = −5.19, t = −2.8), whereas regions more exposed to complementary AI are regaining them (β = +2.37, t = +3.6). The net employment impact is slightly positive once the model is calibrated to match the patterns identified (+0.04 p-points), a sharp deviation from the typical negative impact of robots (−0.37 p-points). This complementarity effect persists after eliminating the influence of pre-existing trends and after controlling for routine task intensity, sectoral concentration, and trade openness.