Even a retail giant like Walmart may not have enough repeat visual tasks to make automating them cheaper than human workers.

This explores whether a company as big as Walmart has enough volume on its own to make automating visual work cheaper than paying people. The short answer from the corpus is no, not for most of it. Svanberg and colleagues built a full cost model of vision automation that counts the expense of collecting data, building and maintaining the system, and running it, then compared that against the wages it would replace. Their result is that even a Walmart-sized firm could cost-effectively automate less than 15% of its vision tasks Can individual firms reach viable scale for vision automation?.

The reason is that building an AI system has large fixed costs, and those costs only pay off when they're spread across enough instances of the same task. Most visual tasks inside a single company, like checking a particular shelf layout or inspecting a particular part, are too narrow or too rare to clear that bar, however big the company is. Across the whole economy, the same model finds that only about 23% of the wages paid for vision tasks are attractive to automate today Which vision tasks actually make economic sense to automate?. The point you might not expect is that "AI can technically do this task" and "AI will actually replace this task" are very different claims. Many headline estimates of AI exposure measure the first, and Svanberg's work suggests the second is much smaller and moves more slowly.

The way past this ceiling isn't getting bigger. It's pooling demand. If many firms share one vision system through AI-as-a-service, the fixed costs get spread across all of them, and tasks that no single company could justify start to make sense Can individual firms reach viable scale for vision automation?. That changes who captures the value of automation. It favors platform providers that serve many customers over the large incumbents you might expect to automate first.

Two other threads in the collection suggest that cost isn't the only filter on what gets automated. Workers are more willing to hand tasks to AI, and to accept less human oversight, when they see those tasks as pointless or disconnected from their organization's goals Does perceived task meaninglessness drive worker desire for AI delegation?. So the tasks that are cheap to automate and the tasks people want automated may not be the same ones. Even when automation does pay off, the gains may not be clean: one model shows that automating entry-level work can raise output while slowing long-run growth, because novices spend less time learning from experts Can automation raise output while slowing growth?.

The corpus is thin beyond Svanberg's study. It doesn't track how much hardware and model costs are falling, or whether foundation models that need less task-specific data are shrinking the minimum viable scale. Those are the main things that could make the 15% figure age quickly.