Workflows and Automation

Labor is typically bundled into jobs, comprising multiple tasks, instead of transacted by the task. This may be because performing a task increases one’s productivity not only at the task itself but at related tasks. I show that these learning spillovers can introduce a convexity to the relationship between automation and output.

Automating the first few tasks in a cluster with high spillovers (a “workflow”) has little or no effect on output, since workers still perform some automatable tasks to increase their productivity at the non-automatable tasks. Automating the remainder of the tasks in such a cluster increases output more than automating the first few, as it becomes efficient to adopt the automation fully. Automation at a larger scale increases output by even more, if machines are intelligent enough also to “learn by doing” and can do a wider range of tasks than any single human.