Introduction

This research explores the fundamental regularity in white-collar occupations where senior professionals, such as lawyers, data scientists, and management consultants, accumulate expertise by executing domain-specific entry-level tasks. In the tradition of economic theory, learning is viewed as a product of experience through activity. The paper highlights a duality in how generative Artificial Intelligence (AI) affects this process. Firstly, a reduction in technology prices may incentivise firms to automate a higher share of entry-level tasks, thereby narrowing the pipeline for skill acquisition. Secondly, the same technology can complement senior managers, helping them expand the task frontier and create new opportunities for junior workers to learn.

Model

The study employs a continuous-time general equilibrium framework where learning-by-doing is determined by three key factors: endogenous automation, an endogenous frontier of tasks, and endogenous career choices. Within the model, atomistic firms do not internalise how their automation and task-creation decisions affect individual human capital accumulation. Managers combine their human capital with technology to maintain a frontier of products or ideas. Households start their careers as workers, accumulating human capital at a rate proportional to the measure of worker-performed tasks, eventually transitioning to management roles at an optimal human-capital threshold.

Characterising Stationary Equilibria

In economies with low learning capacity, the stationary equilibrium is unique, following a standard task-based pattern of no automation at high prices and full automation at low prices. However, in economies with high learning capacity, stationary equilibria exist in pairs. These consist of a high-learning equilibrium with limited automation and lower worker wages, and a low-learning equilibrium characterised by extensive automation or even full automation with zero on-the-job learning. These coexisting equilibria are strictly ranked by the aggregate learning rate, with the high-learning equilibrium providing higher aggregate real value added and greater expected lifetime welfare for households.

A Decline in the Price of the Technology: A Numerical Example

The authors provide numerical solutions to the full model using specific assigned parameter values, including a discount rate of 0.05, a death rate of 0.01, returns to scale of 0.75, and a learning rate of 0.02. The results demonstrate that a decline in technology prices strengthens whichever feedback loop the economy is currently in. In the high-learning equilibrium, cheaper technology reduces the cost of expanding the frontier, accelerating learning. Conversely, in the low-learning equilibrium, cheaper technology makes it more profitable to automate existing tasks, shrinking the learning range and pushing the economy into a ‘human-capital trap’.

Planner’s Problem

The competitive economy is found to be distorted in two directions simultaneously: firms automate too many existing tasks and invest too little in maintaining the task frontier. These distortions arise because atomistic firms fail to internalise the spillovers on the aggregate learning rate. The research suggests that a blanket tax on technology prices would be incorrect, as it would deter both automation and positive frontier expansion. Instead, the first-best allocation is implemented via two distinct instruments: a proportional tax on profits from automation and a proportional subsidy on expenditures dedicated to maintaining the frontier. These instruments take the same rate, determined by a shadow wedge measuring the gap between social and private marginal values of human capital.

Conclusion

The paper concludes that automation and task creation jointly shape career dynamics and long-run human capital. When the automation effect dominates, an economy can fall into a self-reinforcing cycle of reduced learning and lower-quality managers. The planner’s solution corrects this by raising the shadow cost of automation while reducing the shadow cost of frontier expansion. Evaluating automation policy in isolation, without accounting for the role of task creation, will systematically understate the social value of policies that sustain worker learning. Future work could include exploring worker heterogeneity and endogenising technology prices through innovation.