Could pooling AI demand across many companies unlock automation that no single company, however big, can justify alone?

This explores whether pooling demand across many firms, for example by buying AI as a shared service, makes automation pay off where a single company growing bigger can't. The corpus's sharpest answer is yes, at least for some kinds of work. Svanberg and colleagues built a cost model for computer-vision tasks and found that even a Walmart-sized company could only cost-effectively automate less than 15% of the vision tasks it does. The fixed costs of building and running a system are too high to spread over one firm's volume, so growing the firm doesn't fix the problem. Spreading the cost across many firms does, and in practice that means AI sold as a service, where one provider serves everyone's demand (Can individual firms reach viable scale for vision automation?).

The evidence also cuts the other way, which keeps this from being a simple story. On online labor marketplaces, firms more exposed to AI replace contract workers with AI tools faster and more cheaply than less-exposed firms do. That suggests building AI capability inside a firm has its own returns to scale, so adoption is not spreading evenly the way a shared service would predict (Do firms substitute labor for AI at different rates?). One way to fit the two findings together: pooled demand decides whether a capability gets built at all, and firm-level skill decides who gets the most out of it once it exists.

That second step is harder than it looks. Evans argues that making tools cheaper or easier to build doesn't solve the real bottlenecks: most workers don't see their own tasks as automatable, and adoption in a company needs decisions that cross departments and take time (Does easier tool-building actually solve enterprise adoption problems?). Brynjolfsson, Rock, and Syverson add that the investment around AI (retraining, redesigned processes, custom software) is mostly invisible in productivity statistics at first. Early AI adoption therefore looks less productive than it really is, and later it looks more productive than it really is (Why does AI productivity appear weak when investment is high?). A shared service can lower the cost of the model itself, but every firm still has to make these internal investments on its own.

There's also a quieter warning about what 'unlocking automation' optimizes for. One economic model shows that automating entry-level work can raise output while slowing long-run growth, because it pulls junior workers away from the experts who would have passed on hard-to-write-down know-how (Can automation raise output while slowing growth?). If pooled demand makes it cheap to automate routine tasks across a whole industry, that trade-off happens industry-wide, not just inside one company.

The corpus has only one study that directly tests pooling against firm growth, and it covers vision tasks, so treat the strong 'yes' as a well-argued case rather than a settled law. The surprising part is where the bottleneck sits. For many tasks, the barrier to automation is that no single buyer is big enough, so the decisive economic unit may be the market rather than the firm, and AI-as-a-service providers become the ones who decide which tasks get automated first.