Why do companies most affected by AI swap in AI tools faster and for less money than everyone else?

This explores why companies whose work is most exposed to AI replace human labor with AI faster and more cheaply than other companies. The corpus has one direct finding on this. Looking at firms that hire through online labor marketplaces, the more AI-exposed firms swapped those freelancers for AI tools sooner and at lower cost, and each firm did so at its own rate (Do firms substitute labor for AI at different rates?). The explanation offered is returns to scale. Once a firm has built some internal skill at using AI, each later substitution gets cheaper. So AI doesn't spread like a utility that everyone gets at the same price. It spreads more like know-how that builds on itself.

The productivity J-curve helps explain why that know-how matters so much. Brynjolfsson, Rock and Syverson argue that getting value from a general-purpose technology takes large hidden investments: retraining staff, redesigning processes, building internal software. These costs show up long before the payoff does (Why does AI productivity appear weak when investment is high?). A highly exposed firm has more reason to pay those costs early, and once it has, automating the next task costs little extra. Seen this way, "faster and cheaper" may be the visible result of investment made earlier that nobody counted. OpenAI's enterprise telemetry fits the picture: ChatGPT Enterprise adoption concentrates in larger, R&D-heavy firms, which already have the resources to make those investments (Who adopts enterprise AI first and how do they use it?).

There is also a less obvious barrier that exposed firms may have already crossed. Evans argues that the hard part of adoption isn't building tools. It's that most workers don't see their own tasks as automatable, and that company-wide adoption needs decisions spanning many departments (Does easier tool-building actually solve enterprise adoption problems?). A firm whose work is plainly exposed to AI has an easier time with that first step, because the automatable tasks are obvious. Meanwhile the tools themselves keep getting cheaper. Epoch AI finds that the cost of a given level of AI performance drops about 66% per quarter right after it first becomes state of the art (Do AI performance prices fall faster when capabilities debut?). Firms that are already set up to adopt quickly capture more of that early fall in price.

The question of *why* firms choose automation leads somewhere more uncomfortable. Acemoglu, Autor and Johnson argue that firms' profit incentives favor AI that replaces expertise over AI that creates new tasks for workers. No single firm gains from building the pro-worker kind (Why do firms build automating AI instead of pro-worker AI?). If early movers also get cheaper with each step, that bias grows over time. Anthropic's Economic Index offers a partial counterpoint: as adoption matures, usage shifts from handing off whole tasks toward working alongside AI and learning from it (Does AI adoption follow wealth and mature over time?). So the early rush to automate may not be the end state. The longer-run worry is that automation adopted bit by bit wears away the reliance on human workers that keeps institutions answerable to people (Does incremental AI replacement erode human influence over society?). The corpus has only one direct study of firm-level substitution rates, so the self-reinforcing explanation above is a reasonable synthesis, not an established result.