Organizing Intelligence (Google DeepMind, October 2026): teams of complementary humans can beat AI that beats the average worker
Alex Imas, Andrew Koh and Shane Legg of Google DeepMind model the firm, not the single task, as the unit where AI changes the economy. Every worker, human or AI, has a profile of skills, and the paper separates two kinds: competence (how good an agent is at tasks like coding or writing) and coordination (how well agents work together as a firm grows).
Two differences drive their results. AI expertise can be copied for the cost of compute while human expertise is scarce, and AI may not hit the human limits on coordinating large groups. Because people's strengths are uneven in different ways, a team of complementary humans can still outcompete AI that is more capable than the average person, an effect the authors call mosaic intelligence.
When AI gets more competent, firms tend to shrink, with fewer people doing more. When AI gets better at coordinating, it can break the old ceiling on firm size and the market can tip suddenly toward a few large, AI-heavy firms. Job losses are gradual when AI replaces people inside mixed firms, and sudden when AI-native firms outcompete human ones.
For founders, the authors argue that benchmarks of single-task skill are a poor guide to what comes next: a small gain in how AI agents coordinate could reorganize an industry. They are explicit that these are scenarios from the model, not forecasts.