The discourse surrounding artificial intelligence in the labor market has, for the better part of the last decade, generated significantly more heat than light. We stand at a juncture characterized by a pervasive dichotomy: on one side, dystopian determinism suggesting widespread substitution of human labor; on the other, techno-optimist visions of frictionless augmentation. This binary perspective, while rhetorically convenient, is empirically deficient.

It is within this theoretical vacuum that a seminal contribution emerges. In their paper "Roles of Artificial Intelligence in Collaboration with Humans" published in Management Science, Fügener, Walzner, and Gupta^{[Fügener, 2025]} offer a rigorous framework to disentangle the specific economic mechanisms governing human-AI integration. Unlike previous studies treating "AI adoption" as a monolithic variable, they descend into the granular architecture of the task itself.

Their distinction between between-task complementarity and within-task complementarity serves as what I call the "Archimedean point" of their analysis—a fixed reference from which the entire edifice of optimal task allocation can be understood and operationalized.

1The Great Decoupling of Task and Labor

To understand the gravity of this contribution, we must first contextualize it within the broader history of economic thought regarding technological change. Since the industrial revolution, economists have grappled with the "Luddite fallacy"—the fear that machines would permanently displace labor. The standard neoclassical answer has been that while technology substitutes for labor in specific tasks, it complements labor in others, raising productivity and aggregate demand^{[Autor, 2015]}.

However, the AI revolution presents a unique challenge to this equilibrium because it targets cognitive non-routine tasks—the very domain previously identified as the sanctuary of human comparative advantage. Recent work on the "jagged technological frontier"^{[Dell'Acqua, 2023]} suggests that AI capabilities do not scale linearly with task difficulty; rather, they fluctuate unevenly, creating pockets of high performance amidst surprising incompetence.

The fundamental gap in the literature, which this paper addresses, is the lack of a structural model that can predict when AI will substitute for a human (automation) and when it will collaborate with a human (augmentation). Should the radiologist be replaced by the algorithm? Should the algorithm merely advise the radiologist? Or should the radiologist be reallocated to "edge cases" that confound the algorithm entirely?

The answer, as Fügener et al. demonstrate, depends on the precise calibration of complementarity.

2Data and Methodology

To empirically disentangle the effects of automation and augmentation, one requires a data architecture that is both controlled and representative of complex judgment. The authors employ a controlled experimental setting using image classification—a domain where the "Polanyi condition" holds: human decisions rely on tacit knowledge that is difficult to codify into rules.

A critical innovation is the use of the AI's likelihood score (l^AI). For every image, the AI assigns a probability to its classification. This score serves as an ex-ante estimator of task difficulty from the machine's perspective. High l^AI indicates a task that is "easy" for the machine (codified knowledge), while low l^AI indicates a task requiring tacit knowledge.

Using Monte Carlo simulation with 10,000 iterations, the authors model the performance of human crowds. By sampling n humans and aggregating decisions via majority voting, they simulate a "wisdom of crowds" configuration—crucial for the reallocation argument.

3Empirical Findings

The application of the task allocation framework to experimental data yields stylized facts that challenge prevailing narratives of both pure substitution and pure augmentation.

The economic magnitude is profound. Even more striking is the comparison with "Full Augmentation"—the strategy often championed by proponents of "human-centered AI." The optimal framework outperforms universal augmentation by 8 percentage points. This gain is purely a result of allocative efficiency: by automating "easy" tasks, the framework liberates human attention for reinvestment into "hard" tasks via crowd-based mechanisms.

4The Tripartite Structure of Labor

Perhaps the most significant contribution is the empirical mapping of the division of labor. When the optimization model distributes work based on AI certainty (l^AI), a consistent tripartite structure emerges:

5Strategic Implications

This research fundamentally reframes the question facing organizations. The choice is not "automate or augment?" but rather "how do we optimally allocate tasks across a tripartite structure?" This has profound implications for:

This framework also connects directly to recent theoretical work on the knowledge economy^{[Ide, 2025]} and organizational hierarchies^{[Garicano, 2000]}. The tripartite structure mirrors the stratification of cognitive labor: routine tasks to machines, complex judgment to human-AI teams, and irreducible tacit knowledge to aggregated human expertise.

6Conclusion: Beyond the Binary

The Fügener, Walzner, and Gupta framework provides the Archimedean point we have been seeking: a fixed reference from which to understand the complex dynamics of human-AI collaboration. Their work demonstrates that the "future of work" is not a monolithic shift but a stratified ecosystem, where the division of labor is dictated by the limits of codification and the persistence of tacit knowledge.

For practitioners, this means moving beyond the "AI everywhere" heuristic. For researchers, it opens new questions about how complementarity structures evolve as AI capabilities improve. And for policymakers, it suggests that labor market impacts will be highly heterogeneous—varying not just by occupation, but by the granular task composition within each role.