Compression inside, divergence outside
The surprise finding
Brynjolfsson, Li, and Raymond's 2023 NBER study of customer service agents using a generative AI assistant found a striking pattern. The most experienced agents saw little productivity boost. The least experienced and most junior agents saw dramatic gains — often 30 to 40 percent. The AI was not substituting for the top of the skill distribution; it was bringing the bottom up toward the top.
Compression inside firms
This within-firm compression challenges old assumptions. In previous automation waves, machines tended to widen skill premiums by rewarding the scarce humans who could operate them. AI assistants often do the opposite at the task level, at least at the early-career end: they codify senior judgment and make it available to juniors. Seniority therefore commands a smaller premium than it used to, inside firms adopting these tools.
Divergence between firms
The same research shows the opposite pattern at the firm level. Firms that adopt AI early and well capture disproportionate gains in productivity, revenue, and market share. Firms that do not adopt fall behind. The result is compression within each adopting firm and widening gaps between adopters and non-adopters.
What this means for inequality
The inequality story is therefore more complex than 'AI helps the top, hurts the bottom'. Inside a firm, adoption can compress pay. Across firms, adoption can widen it. The net effect depends on whether you are in an adopting firm, an adopting industry, or an adopting country — and on whether capital ownership of those firms is broad or narrow. Piketty's work on capital concentration is directly relevant here.
If an AI tool made the worst performer on a team 40% better and left the best performer unchanged, what should happen to pay?