Which vision tasks actually make economic sense to automate?

Svanberg, Li, Fleming, Goehring and Thompson, in an MIT FutureTech working paper, build what they call "the first end-to-end AI automation model" to separate whether a task is technically possible to automate from whether automating it is worth the money. Restricting the analysis to computer vision, where cost modeling is most developed, they report that "at today's costs U.S. businesses would choose not to automate most vision tasks that have 'AI Exposure,'" and that "only 23% of worker wages being paid for vision tasks would be attractive to automate."

The mechanism is a concept they name minimum viable scale: the point at which an AI deployment's fixed costs are amortized enough that its average cost per task matches the cost of human labor doing the same task. Below that scale, keeping the human is cheaper even when the AI system can technically do the work. The authors model human labor cost as straightforward labor accounting and model AI system cost from "the computer science literature on training and doing inference with deep learning," supplemented by 35 case studies they conducted themselves; automation is only cost-effective, in their framework, "when the deployment scale is larger than the minimum viable scale."

This cuts against the prior "AI Exposure" literature (Felten et al., Eloundou et al.'s widely cited estimate that roughly 50% of tasks could be partially automated by LLMs), which the authors say is "vague about the timeline and extent of automation because [it does] not directly consider the technical feasibility or economic viability of AI systems." It also complements Do firms substitute labor for AI at different rates? and Does concentrated AI exposure enable workers to adapt and reallocate?: both of those findings describe firm-level heterogeneity in AI adoption speed, and this paper supplies a cost-side mechanism — scale versus the minimum viable threshold — for why exposure alone does not predict adoption. The companion note Can individual firms reach viable scale for vision automation? extends this same mechanism to the firm-size question directly.

The finding is scoped to computer vision tasks and to "today's costs" as the paper's 2024-01-18 dataset recorded them — it does not model LLM-based automation directly, though the authors argue in their discussion section that the framework generalizes. It also does not estimate how quickly costs will fall or how fast AI-as-a-service aggregation will spread, only that either would accelerate the 23% figure upward. The authors' own policy conclusion follows at the strength their evidence allows: "AI job displacement will be substantial, but also gradual — and therefore there is room for policy and retraining to mitigate unemployment impacts."