AI and Scale: A Quantitative Task-Based Theory of Automation

Source: Danial Lashkari, Christina Qiu, Wensu Li and Neil Thompson (MIT FutureTech); working paper, PDF titled 'AI, Scale, and Task-Based Theories of Automation' · 2026-07-13

AI automation exemplifies technologies that entail task-level fixed costs, e.g., model training or fine-tuning. We integrate such fixed costs into a general theory of automation and show that AI’s comparative advantage becomes scale-dependent at the task level. We derive the elasticities characterizing the resulting production function, including the endogenous labor–AI substitution elasticity. We apply a quantitative version of the model to computer vision automation, disciplining fixed (training) and marginal (inference) costs using AI scaling laws, estimated from a fine-tuning experiment. Calibrated to 2023 U.S. data and recent estimates of the decline in compute prices, the model projects rapid automation, reaching 23% of firms (60% of employment) by 2035. The labor share traces a U-shape as the substitution elasticity declines from above to below one. Scale accounts for the majority (∼63%) of cross-task variation in AI comparative advantage.

Introduction. The recent surge in the capabilities of Artificial Intelligence (AI) systems has turned the labor-market consequences of automation into an urgent question for economic research.

A growing body of empirical (e.g., Brynjolfsson and Mitchell, 2017; Felten et al., 2018; Frey and Osborne, 2017) and theoretical (e.g., Acemoglu and Restrepo, 2018a, 2020) work studies this question at the level of the tasks required to fulfill production. Through this lens, the relevant question is whether machines (AI) have a comparative advantage over labor in performing any given task, in the sense of offering lower production costs. In canonical task-based models, this comparative advantage is governed by relative marginal costs, which, in the case of AI, are primarily driven by the computational costs of running the model to perform the task (i.e., the costs of inference). Yet this standard approach abstracts from the defining role of scale, both in driving progress in AI performance and in deploying AI in production. AI scaling laws establish a robust empirical relationship between AI performance and the computational inputs used to train models (Hoffmann et al., 2022; Kaplan et al., 2020).1 Since the corresponding training costs are fixed with respect to the volume of inference, a larger scale of AI output justifies developing a more capable model. This logic also applies at the task level since models require fine-tuning to acquire capabilities specific to a given production task (Fleming et al., 2024). Consider two tasks that are both exposed to automation by AI: detecting tumors in radiological images, which has so far resisted automation, and monitoring a retailer’s in-store cameras.2 Automating either task requires its own model, fine-tuned on data collected and labeled from that setting. Whether such a task is worth automating then turns not only on whether machines are cheaper than labor at the margin, but also on whether the task is performed at enough scale to spread this fixed cost. Production scale may thus be a source of AI comparative advantage across tasks, alongside the marginal-cost comparison that standard theories emphasize. Motivated by this observation, we provide a general theory of task-based automation, in which automating a task involves, in addition to variable (marginal) costs, a fixed cost of adopting machines. In this environment, automation is governed by scale-dependent comparative advantage, the product of two terms: the standard marginal advantage, i.e., the marginal productivity of machines relative to labor, and a novel scale advantage, an increasing function of the ratio of variable costs to fixed costs and, therefore, of the task scale. As task output grows or fixed costs vanish, scale advantage approaches unity and Page 3 Section 4 uses these estimates of AI’s cost structure to quantify the model. We calibrate the remaining parameters using standard values from the literature, together with data on firm-size distributions and on firms’ adoption of computer vision AI in 2023 (Bonney et al., 2024). The calibrated model underlies the results described above, which we further illustrate through the automation path of a single firm, where adoption arrives in waves across groups of tasks with similar complexity and error tolerance. We also show similar results when the fixed cost of fine-tuning is incurred at the industry rather than the firm level, and under an alternative scenario featuring a faster compute-cost decline that incorporates algorithmic progress.

Related work. Our paper contributes to task-based theories of automation (Zeira, 1998; Autor et al., 2003; Acemoglu and Autor, 2011; Acemoglu and Restrepo, 2018b, 2019) in three ways. First, we incorporate a central feature of AI automation: task-level fixed costs that arise when a general-purpose foundation model must be specialized to a particular production task, for instance, through fine-tuning. Second, for this general model, we characterize the dual cost function and the full set of production elasticities it implies, including the labor–AI elasticity of substitution and the output elasticities of factor demand. We show that these objects are scale-dependent in the presence of fixed costs and evolve endogenously as machine prices fall. Third, we specialize the general theory to a quantitative model that directly maps to task classifications such as ONET. The mapping draws on tools commonly used in quantitative trade. Viewed through this lens, our framework introduces fixed costs and scale dependence into canonical models of offshoring (Grossman and Rossi-Hansberg, 2008) and comparative advantage (Dornbusch et al., 1977), with production tasks playing the role of traded commodities. Although AI’s potential for automation is widely recognized, direct evidence on its productivity effects remains largely confined to narrow experimental settings (e.g., Dell’Acqua et al., 2023; Peng et al., 2023; Cui et al., 2024; Brynjolfsson et al., 2025). A broader empirical literature instead measures task or occupational exposure to AI. Following early contributions such as Brynjolfsson and Mitchell (2017) and Felten et al. (2018), this literature constructs proxies for the extent to which AI can affect production across tasks and combines them with data sources such as ONET to measure occupational exposure (see also Brynjolfsson et al., 2018; Frank et al., 2019; Felten et al., 2021; Lane and Saint-Martin, 2021; Cazzaniga et al., 2024).5 These measures capture only whether AI capabilities are relevant for a task, not whether AI can meet the task’s performance requirement, the magnitude of task-level productivity gains, or the costs that govern adoption decisions. Our framework instead turns task characteristics into cost primitives and incorporates the fixed costs of adoption, tracing their implications for the allocation of production across firms and industries and, ultimately, for aggregate outcomes. A central methodological ingredient in our quantitative exercise is the use of AI scaling laws to infer how training and inference costs vary across tasks. An alternative is to infer comparative advantage from observed adoption patterns as such data become available (e.g., McElheran et al., 2024; Bonney et al., 2024).

Method. We begin in Section 2.1 with a general model of task-based production in which using machines requires task-level fixed costs, and we characterize the firm’s automation decisions and cost structure. Section 2.2 introduces a task-type representation that preserves the generality of this model while allowing us to derive the key elasticities characterizing the production function and the factor demand system. Section 2.3 then specializes this representation for quantitative analysis, illustrates its implications in two limiting cases, and embeds firms in a general-equilibrium environment that maps directly to task-level data.

In this section, we introduce a representation for the model of Section 2.1 that, while maintaining the model’s generality, allows us to examine the properties of the resulting production function. More specifically, under this representation, we derive the key elasticities that characterize the production function and the resulting factor demand system, including the elasticity of substitution between labor and machines and the output elasticity of demand for machines.

Equation (21) is one of the key results of our paper. It shows that the firm’s labor demand elasticity always exceeds the task elasticity of substitution λ, which accounts for the standard intensive margin of substitution in a CES production function without the extensive margin. Beyond this margin, a rise in the price of machines r further raises demand for labor since automation cutoffs for all task types fall due to the decline in machine comparative advantage along two dimensions. First, a rise in machine price r brings about a decline in marginal advantage, which is present in the standard models of task-based automation. This term, captured by the second term on the right-hand side of Equation (21), is higher when workers spend more time on task types with a higher local elasticity of the selection function, as captured by the expression EL h eGλ−1 (z∗ τ|τ) i . This elasticity gives the percentage decline in the labor cost shifter φL,τ per percentage rise in the automation cutoff z∗ τ, capturing how strongly the labor component of production responds to automation’s extensive margin. Second, in the presence of fixed costs, a rise in machine price r brings about an additional decline in the scale advantage of machines, since the task scale yK(ν) also falls across all tasks.

Discussion. Next, we assess the quantitative importance of scale advantage in shaping comparative advantage across tasks. From Equation (5), for a task item τ, comparative advantage follows log CAτ = log MAτ + log SAτ. This logadditive structure allows us to decompose the variance of comparative advantage into scale- and marginal-advantage components by regressing each component on log comparative advantage.21 Figure 3 displays the results of this exercise for the calibrated baseline: in 2023, the scale component accounts for 63% of the cross-task variation in comparative advantage. This finding is robust to performing the decomposition using averages across broader task categories, as shown in Appendix Figure A5. Thus, most variation in comparative advantage arises from scale rather than marginal cost differences, indicating that scale considerations are a crucial determinant of task-level automation patterns. Proposition 1 implies that comparative advantage eventually converges to marginal advantage as aggregate output grows. Quantitatively, however, this asymptotic result has limited relevance over economically meaningful horizons. When the model is projected forward to 2050, scale advantage still accounts for roughly 60% of the cross-task variation in comparative advantage. Appendix Figure A6 reports this variance decomposition. Figure 4 shows the average of AI comparative advantage and its components within selected bundles of task groups at our baseline calibration.22 Here, AI exhibits the greatest comparative advantage in “detecting unusual behavior or visual signals in operational settings” and the least comparative advantage in “analyzing medical imaging”.

We use the calibrated model to project the path of computer vision AI automation from 2023 to 2075 as computing costs decline. Figure 5 shows rapid growth in adoption over the first decade of the transition, especially when firms are weighted by employment. By 2025, the unweighted adoption rate reaches 4.7%, while the employment-weighted adoption rate reaches 12.2%. By 2035, these rates rise to 22.8% and 59.5%, respectively. These adoption rates capture the extensive margin of firm-level automation: the share Page 33 of firms for which more than 10% of vision-task varieties are automated. To measure the intensive margin, Figure 5 also reports the cost-weighted average share of automated task varieties across all tasks and firms in the economy, Ej[Eτ|j[φK,ijτ]]. Because the 659 vision tasks account for only a small share of all production tasks, this task-level adoption measure only reaches 6.4% over the projected automation path. Throughout the transition, employment-weighted adoption remains above unweighted adoption, indicating that larger, more productive firms adopt computer vision AI earlier and more extensively.23,24 Our quantitative framework allows us to project the outcomes of specific firms along the automation path, including their AI expenditure and firm-level labor demand elasticities of substitution. To illustrate the rich predictions of our quantitative framework, we next focus on the outcomes of a firm in the industry “Gasoline Stations” (NAICS 4571). Figure 6 shows the model-implied automation path for a representative firm in this industry. The path mirrors the aggregate pattern in Figure 5, with automation expanding as compute costs decline.

Conclusion. This paper develops a task-based theory of automation in which adopting machines entails task-level fixed costs. Fixed costs make machine comparative advantage scaledependent: even when machines are cheap at the margin, automation depends on whether task scale is large enough to amortize the fixed cost. The canonical model appears as the limiting case in which fixed costs vanish or task scale becomes large. The paper also proposes a task-type representation that bridges canonical theories of automation to their quantitative implementations. Using this representation, we can derive the endogenous elasticities of substitution between factors and output elasticities of factor demand. We use the representation to study computer vision automation, relying on AI scaling laws to discipline AI’s marginal (inference) and fixed (training) compute costs across tasks. In our application, scale explains the majority of cross-task variation in comparative advantage, and its importance remains stable when projecting the economy forward along a path of declining compute prices. Along this path, the aggregate labor– AI elasticity of substitution declines from above to below one, producing a U-shaped labor share even as output and wages rise. We also demonstrate that scale remains important when assuming that fixed costs are sunk at the level of the industry, instead of the firm. Future work can examine how intermediate cases—where, for instance, firms outsource the fine-tuning process to upstream providers—may lead to different patterns of automation. We emphasize that the theoretical results of the paper generalize to a variety of contexts beyond computer vision AI where fixed costs may shape comparative advantage, including other automation domains and applications in outsourcing, offshoring, and commodity trade. Our task-type representation can likewise quantify task-based models of automation for other technologies, such as LLMs or robotics—even ones where fixed costs play a minor role—provided credible task-level measures of their cost structure are available.