Even Walmart-scale operations can't automate most visual-inspection jobs profitably — so what's actually blocking AI adoption here?

This explores how big an operation has to be before replacing human visual work (inspecting parts, reading images, checking shelves) with AI actually beats paying people, and whether any single company is big enough. The corpus's answer is blunt: almost none are. In Svanberg and colleagues' cost model, only about 23% of the wages paid for vision tasks are economically attractive to automate today, even though AI could technically do far more of them Which vision tasks actually make economic sense to automate?. The gap between "AI can do this" and "it's worth building AI to do this" turns out to be the main story.

The surprise is how high the bar sits. The same model finds that even a firm the size of Walmart could cost-effectively automate less than 15% of its vision tasks on its own Can individual firms reach viable scale for vision automation?. So the minimum viable scale isn't a headcount or revenue threshold a company can grow into. For most vision tasks it sits beyond what any single firm reaches. The suggested way around this is pooling across companies: AI-as-a-service providers spread the upfront cost of building and fine-tuning a system over many customers until volume is large enough.

The reason scale matters so much is fixed costs. Lashkari and colleagues show that the one-time cost of adapting a model to a specific task explains about 63% of the variation in whether automating that task pays off Does AI's automation advantage depend on how often a task runs?. Two equally hard tasks can land on opposite sides of the line purely because one runs millions of times and the other runs thousands. That means the useful question isn't "how big is my firm?" It's "how many times does this exact task run, across everyone who does it?"

That threshold is moving. Epoch AI's Denain estimates that the cost of reaching a given capability level falls roughly 5–10x a year Does inference cost remain a persistent economic burden?, and work on compact models suggests smaller, task-trained systems can match much larger ones at a fraction of the cost Does model efficiency matter more than peak capability for real work?. Neither addresses vision directly, so treat this as a reasonable inference rather than a finding: cheaper adaptation and cheaper running costs should pull more tasks below the viability line over time. The corpus doesn't give a dated forecast for vision specifically.

Here's a less obvious point. Even after a task becomes cheap enough to automate, the value doesn't simply leave the job. Gans and Goldfarb argue that when tasks depend on each other, automating some of them makes the remaining human steps more valuable Does automating some tasks actually increase the value of remaining work?. Other work suggests the next bottleneck becomes the human capacity to check what the automated system produced What becomes scarce when AI execution grows cheap?. So the economics of vision automation may come to depend less on whether the system is cheap enough and more on whether anyone can afford to verify its output.