Where have workers actually delegated tasks to AI?

The paper adds a tier to the AI-exposure literature. Existing measures "capture where AI could perform tasks rather than whether workers have actually adopted it," so the authors define delegated exposure as whether a worker "has committed a task to AI by embedding it into a structured workflow." Their Agentic Adoption Index (AAI) scores how closely an occupation's tasks match agentic routines that practitioners have built and shared. The reported gradient is that delegated exposure "concentrates in information-intensive work and remains minimal where work depends on manual dexterity or direct intervention in the physical environment." It also departs from pre-AI automation frameworks: the occupations where delegation concentrates "differ sharply" from those flagged as most vulnerable, and computerization-risk estimates, which located risk in routine tasks, ranked occupations differently.

The measurement rests on a change in how LLMs get used. As use moves "beyond isolated conversations toward reusable agentic workflows," the reusable skill specification becomes the observable unit. The authors embed roughly 888,000 skill specifications from public GitHub repositories, compute their similarity to nearly 18,000 O*NET task statements, and aggregate to the occupational level. The second finding complicates the "adoption" framing. The AAI aligns more closely with measures of technical capability and application availability than with prompt-level conversational use, which the authors read as suggesting it "tracks what AI can in principle perform rather than what usage records currently capture." Availability is not the whole story either. Adoption rises with wages among occupations requiring a bachelor's degree or less, but declines among higher earners requiring advanced degrees.

Against the nearest notes, this paper contributes a different kind of exposure evidence. Does concentrated AI exposure enable workers to adapt and reallocate? works with task-level exposure and its concentration within occupations, while the AAI asks which tasks have already been written into workflows. Do firms substitute labor for AI at different rates? observes behavior at the firm level, whereas the AAI is occupation-level and drawn from public repositories. The premise that value migrates from conversation to persistent, reusable routines is the labor-measurement counterpart of What makes an AI system feel like a colleague rather than a chatbot?. The AAI's high-wage decline is also a possible point of contact with What collaboration level do workers actually want with AI?, though the excerpt draws no such link.

The excerpt does not establish how the AAI was validated, how strong its alignment with capability or usage measures is, or which occupations sit at either end of the gradient. It states no effect sizes. It says nothing about who wrote the public skill specifications, so it cannot show that workers in the matched occupations are the ones delegating. It also says nothing about employment outcomes. The abstract promises three main findings but the excerpt gives two, and the wage pattern appears only in the discussion, with no explanation offered for the decline among advanced-degree occupations. The safe reading is narrower than the title: public agent-skill sharing clusters in information work and follows capability more than chat logs do. Whether that counts as realized adoption is left open.