The cost of AI accuracy is convex. Getting from 80 percent to 90 percent accuracy is affordable. Getting from 90 to 95 costs considerably more. Getting from 95 to 99 costs enormously more. Getting from 99 to 99.9 may cost more than keeping a human in the loop. Li and colleagues formalize this through AI scaling laws — the relationship between performance and the data, compute, and model size required to achieve it.

The implication: partial automation isn’t a transition phase on the way to full automation. For many tasks, it’s the long-run economic equilibrium. The cost curve bends upward fast enough that the marginal dollar spent on accuracy eventually exceeds the cost of human oversight. At that inflection point, the rational decision is to stop automating and keep the human for the last few percent.

The entropy-based task complexity measure sharpens this. Simple tasks — low entropy, predictable outcomes — have cheap accuracy curves and high automation potential. Complex tasks — high entropy, context-dependent outcomes — have steep accuracy curves where near-perfect performance is prohibitively expensive. Computer vision captures roughly 11 percent of exposed labor compensation at the firm level. Not 90 percent. Not 50 percent. Eleven — because the convex ceiling on most tasks arrives before full displacement.

The structural lesson: scaling laws aren’t just about capability. They’re about cost. The same scaling law that predicts an AI will eventually reach human performance also predicts how expensive that last mile will be. The performance trajectory and the economic viability are different curves. One keeps rising. The other bends. Where they diverge is where partial automation stabilizes — and that’s the equilibrium, not the transition.