If a company gets good at using AI, does it keep getting cheaper and better at it — pulling ahead of rivals using the same tools?

This explores whether AI capability builds up inside firms, so that early, heavy adopters pull further ahead instead of every company getting the same boost from the same tools. The corpus has one direct finding on this. The rest of the evidence fits with it, but it comes from the side rather than testing the question head-on. The direct finding: firms with more AI exposure replace freelance-marketplace workers with AI faster and at lower cost than less-exposed firms. That points to firm-specific returns to scale rather than a technology that diffuses evenly Do firms substitute labor for AI at different rates?. In plain terms, the second, third and tenth AI deployment seem to get cheaper for firms that have already done the first.

Who adopts first fits the same pattern. OpenAI's enterprise telemetry across 1,764 firms shows ChatGPT Enterprise use concentrated in larger, R&D-intensive companies. Inside those firms, use is uneven: marketing staff and early-career workers use it far more than executives Who adopts enterprise AI first and how do they use it?. The firms best placed to build compounding capability are already big and already invest in experimentation. A survey of 750 executives found that gains cluster in high-skill services and finance, and that perceived productivity gains run ahead of measured ones, partly because revenue lags behind operational improvements Do AI productivity gains feel larger than they actually measure?. If the returns build quietly before they show up in revenue, the gap between firms may be wider than the numbers currently show.

The less obvious part is what the scale advantage is made of. It's probably not access to models, which anyone can buy. Evans argues that making AI tools easier to build doesn't solve the real bottlenecks. Workers often don't see their own tasks as automatable, and adoption needs decisions that cut across departments and budgets Does easier tool-building actually solve enterprise adoption problems?. That kind of organizational know-how is exactly what accumulates. A firm that has learned to spot automatable work, and to push changes through its own structure, does it faster next time. The same logic shows up in evaluation. Agents that win benchmark contests still fail at long, multi-step professional workflows Why do agent benchmarks not predict real economic value?. And in real multi-step work, cost and latency pile up over a whole task, so efficient deployment counts as much as raw model power Does model efficiency matter more than peak capability for real work?. Closing that gap from model to actual work is internal engineering and process work, and firms get better at it with practice.

There's also a cost worth noticing. If capability compounds through replacing labor, the firms that scale fastest also become the least dependent on human workers who care about outcomes. The gradual-disempowerment argument holds that this dependence is one of the quiet ways institutions stay aligned with human preferences, and that removing it piece by piece can cause drift that's hard to reverse Does incremental AI replacement erode human influence over society?. Returns to scale and loss of human influence may be two views of the same process. The corpus doesn't yet have firm-level panel studies, or studies of whether late adopters catch up. So treat "yes, probably, and mostly through organizational learning" as a well-supported hunch, not a settled result.