Does AI's automation advantage depend on how often a task runs?

Danial Lashkari, Christina Qiu, Wensu Li and Neil Thompson, in an MIT FutureTech working paper, argue that "AI automation exemplifies technologies that entail task-level fixed costs, e.g., model training or fine-tuning," and that incorporating those fixed costs into a general theory of automation shows "AI's comparative advantage becomes scale-dependent at the task level." They decompose a task's comparative advantage into two multiplying terms, writing "log CAτ = log MAτ + log SAτ": a standard marginal advantage (machines' marginal productivity relative to labor) and a new scale advantage, "an increasing function of the ratio of variable costs to fixed costs." Calibrated to 2023 U.S. data and computer vision automation specifically, the paper reports that "the scale component accounts for 63% of the cross-task variation in comparative advantage" — more than marginal-cost differences.

The mechanism is that a model must be fine-tuned on data specific to a given production task before it can perform that task, and this fine-tuning cost is fixed regardless of how much the task is ultimately run. Whether automating a task pays off then depends on "whether the task is performed at enough scale to spread this fixed cost," not only on whether the AI is cheaper than labor at the margin — the paper's own example contrasts detecting tumors in radiological images, "which has so far resisted automation," with monitoring a retailer's in-store cameras, a higher-volume task with comparable technical exposure. As task scale grows large or fixed costs vanish, scale advantage converges toward the canonical marginal-cost-only model, but the authors find this convergence has "limited relevance over economically meaningful horizons": projected to 2050, scale still accounts for "roughly 60%" of cross-task variation.

The paper shares its minimum-viable-scale logic with Which vision tasks actually make economic sense to automate? and Can individual firms reach viable scale for vision automation? — both MIT FutureTech papers with overlapping authors (Wensu Li and Thompson appear on all three) — but formalizes the mechanism into continuous elasticities and a general-equilibrium calibration rather than firm case studies. Projected forward, the calibrated model has computer vision adoption reaching an unweighted firm rate of 4.7% by 2025 and 22.8% by 2035, with employment-weighted rates of 12.2% and 59.5% respectively, while the economy-wide labor share "traces a U-shape" as the labor-AI elasticity of substitution declines from above one to below one. This sits alongside How much will AI actually boost productivity over ten years? as a second quantitative task-based automation model, though this one's bound comes from scale economics rather than that note's aggregation method.

The paper is a working paper, not yet peer-reviewed, and its quantitative results are calibrated to a single domain — computer vision — using AI scaling laws and a fine-tuning experiment to estimate cost parameters, then projected forward to 2075 under an assumed path of declining compute prices; the 2035 and later figures are model projections, not observed adoption. The authors state their task-type representation "can likewise quantify task-based models of automation for other technologies, such as LLMs or robotics," but only "provided credible task-level measures of their cost structure are available" — a condition the excerpt does not show has been met for any technology beyond computer vision.