How do cognitive tools reshape the three types of worker skill?

In a July 2025 NBER working paper (w34034, not yet peer-reviewed), Ajay Agrawal, Joshua Gans and Avi Goldfarb formalize Steve Jobs's "bicycles for the mind" metaphor into a model of cognitive tools — computers, AI prediction, and generative AI — that improve "the ratio of success probability to cost for any given level of effort" in iterative task improvement. They decompose human skill into three components: implementation skill (performing the task), "opportunity judgment" (the ability to recognise opportunities to start or improve a process), and "payoff judgment" (knowing "which action to take in a given state"). The paper's central claim is that these three skills relate to cognitive tools in structurally different ways: tools are substitutes for implementation skill, opportunity judgment "is shown to always complement cognitive tools," and payoff judgment is "not necessarily a complement" — it complements tools only "as long as the reduction in effort is not so large as to reduce output."

The mechanism: adopting a cognitive tool lets an agent reduce direct effort because the tool raises the odds of success per unit of cost, but because the net benefit from each improvement rises, judgment about when and how to use that freed-up effort becomes more valuable rather than less. The authors illustrate with CAD software, which amplifies opportunity judgment by letting a designer act faster on spotted improvements, and with AI diagnostic prediction, where payoff judgment translates a predicted state into the right treatment decision — a complement to the tool only so long as the effort reduction doesn't outweigh the accuracy gain.

This framework is explicitly offered to reconcile contradictory empirical findings: "computers increased labour market inequality... while AI often reduces inequality within adopting firms." The authors attribute computer-era inequality to high variance in opportunity judgment relative to implementation skill, and AI's current equalizing effect to its substituting for implementation skill that less-experienced workers lack. The paper states "our approach complements Autor and Thompson (2025)," directly extending the logic behind Does automation raise or lower the skills that remaining work demands? by supplying a skill-level mechanism underneath Autor and Thompson's task-level one. It also sits alongside Why do firms build automating AI instead of pro-worker AI? as a second attempt to classify, rather than merely measure, how AI's design interacts with worker value, and alongside What makes accountable judgment scarce when AI cognition is cheap?, which treats judgment as the scarce resource without this paper's split between opportunity- and payoff-type judgment. Does better prediction change which human judgment matters most?, a companion paper by the same authors, develops a related judgment distinction in the context of prediction specifically.

The paper is a formal theoretical model, not an empirical test: it is built to interpret existing studies it cites (Brynjolfsson et al. 2025, Cui et al. 2025, Autor et al. 2006, among others) rather than to produce new measurements, and as an NBER working paper it has not been peer-reviewed. Because opportunity and payoff judgment are modeled as latent skills rather than measured variables, the paper does not establish how to identify or quantify either skill in a real workforce — only that, if the model's assumptions hold, their distribution relative to implementation skill determines whether a cognitive tool raises or lowers inequality.