Does automating some tasks actually increase the value of remaining work?

Gans and Goldfarb model a job as an "O-ring technology," borrowing Kremer's 1993 term for a production function where "output is the product of task qualities" rather than a sum of them. Because qualities multiply, "a change in the quality of one task scales the marginal value of quality in every other task." Automating a task is therefore not a clean subtraction of labor demand: it is, in the authors' words, "a reallocation event." The paper's headline implication is that "widely-used exposure indices, which aggregate task-level automation risk using linear formulas, will overstate displacement when tasks are complements." The object that matters is "not average task exposure but the structure of bottlenecks and how automation reshapes worker time around them."

The mechanism is what the authors call a "focus" effect. A worker has a fixed time endowment h to split across whichever tasks remain manual; when a machine takes over some tasks, "the worker's fixed time endowment is concentrated onto the remaining manual tasks, endogenously raising their quality." Human quality on a task is modeled as a(h/m), where m is the shrinking count of manual tasks — "an endogenous equilibrium object," not the fixed baseline that exposure indices (Frey and Osborne 2017, Webb 2020, Felten et al. 2021) assume when they compare machine quality against a static human quality. Two further results follow from solving the firm's choice of which tasks to automate under Nash bargaining: automation decisions are "discrete and bundle-like" rather than smooth and marginal, and labor income "need not fall under partial automation" because automation can scale up the value the worker captures from remaining bottleneck tasks. A task that "looks automatable today... may become non-automatable tomorrow."

This sits alongside Does automation raise or lower the skills that remaining work demands? as a second, independent reason simple task-removal accounting fails: Autor and Thompson's channel is which tasks leave an occupation's bundle and how that shifts the expertise demanded of the rest, while Gans and Goldfarb's channel is the multiplicative production structure itself, operating even without any change in required expertise. Why do firms build automating AI instead of pro-worker AI? sorts automation by its effect on worker value from the outside; the O-ring model supplies a mechanism for why that effect can run positive under "task-replacing" automation specifically when tasks are complements, not just when new tasks are created. It likewise bears on How much will AI actually boost productivity over ten years? and on Does AI exposure hit low-wage workers harder in some fields?, both of which build on task-exposure measures the paper argues are biased toward overstating displacement wherever task complementarity holds.

The model is "deliberately spare": a single worker, one firm, Nash bargaining, and no empirical calibration to real occupations, AI systems, or exposure-index data. The excerpt names no estimate of how much existing indices overstate displacement, and it does not say which real jobs look O-ring versus separable — that classification is left as an open empirical question (Section 6 is explicitly about "prospective and retrospective task-based assessments," not new measurement). The implication the paper supports, at the strength a theoretical result allows, is that any exposure-index estimate of AI displacement should be read as an upper bound wherever the job's tasks are quality complements rather than separable.