DeepMind Breaks Down the AGI Economics: Even the Strongest Single Model Can Lose to a Human Team

Google DeepMind's Alex Imas, Andrew Koh, and Shane Legg have published a study titled The Economics of Intelligence. The accompanying paper is Organizing Intelligence. DMI is a platform founded by Google and DeepMind researchers; a note at the end states that the article does not represent Google's official position.

They ask a direct question: AI capabilities are rising fast, so why does the macroeconomy seem barely changed? Their answer up front: comparing a single AI agent to a single person is the wrong unit of economic analysis. Firms are not piles of average people; they are organizations that combine different strengths. Economic transformation is the equilibrium outcome of organizational competition.

Original paper: Organizing Intelligence

DMI article: The Economics of Intelligence

They first ran a small experiment. 100 people took a test covering six domains: math, verbal reasoning, science, patterns, culture, and economics. Being strong in one domain did not predict being strong in another. There is no copy of average intelligence in the population, only specialists oriented in different directions. Organizations combine complementary people into mosaic intelligence, so the team's capabilities cover individual gaps.

In the experiment, pairing complementary people to solve specific tasks performed 32% better than random pairing. This is not because any individual became stronger, but because the organization stacked each person's comparative advantage together.

AI has the same problem. Models are jagged too, but they are jagged in similar ways. Copying a model adds capacity, not breadth of capability. So a model beating the average person does not mean it can beat a complementary human team. Asking whether a model is stronger than a person is the wrong question. The right question is whether AI organizations can outperform human organizations, or hybrid human-AI organizations.

They split economic impact into two thresholds. Competence determines the level of output per unit, and coordination determines how large an organization can scale. Organizations also have scaling laws, in three stages: solo work, bounded scale, and unbounded scale. Past the supercooperative threshold, each additional member makes the organization more productive, breaking the Coase ceiling. Firm size is no longer capped by coordination costs.

So there are four regimes.

Minimal AGI: replicable AI reaches or exceeds average human capability. Structural change is small; wages move first. Wages for skills matched by AI move toward compute costs. Experts and coordinators may see higher wages. Small AI companies enter price-sensitive and niche markets.

Supercompetent AI: AI exceeds human experts on almost every individual capability, but coordination remains constrained. The result is counterintuitive: abundance without concentration. Firms are more efficient but not larger; headcount may shrink; many lean companies compete; goods and services become cheaper. Human comparative advantage shifts toward judgment, orchestration, and designing coordination architectures.

Supercooperative AI: individual capability stays at the minimal AGI level, but coordination crosses the threshold. The economy suddenly undergoes a phase transition. Highly synchronized AI organizations in deep markets rapidly grow large, defeat small companies, and cause abrupt changes in market concentration, employment, and wages. Wages in deep markets plummet. Niche markets have limited demand, but labor inflow pushes wages down.

ASI: both thresholds are crossed. Deep markets concentrate in large AI companies, and niches no longer shelter human teams. The relational sector may expand because human participation is itself valuable.

Today the economy has not changed much because the competence index is rising while coordination lags. OpenAI's Noam Brown said on the Dwarkesh podcast that running thousands of agents in parallel to solve open math problems makes it hard for them to coordinate effectively. They tend to work independently rather than divide labor. He said he would not attribute more than 10% of the credit to multi-agent systems.

One commenter suggested directly measuring a coordination tax: fix single-model capability, scale an AI team from 1 to 5, 20, and 100, and measure output quality, cost, duplicated work, handoff failures, human intervention, and completion time to see whether marginal coordination cost falls. That suggestion is closer to the real question than another single-agent leaderboard.

The conclusion is not complicated. Wages and job roles will move earlier than GDP or industry concentration. We need organization-level benchmarks that track payroll and compute spending. What is really worth watching is whether agent teams can divide labor, hand off cleanly, and correct errors over time. Otherwise, no matter how high a single model's benchmark score is, it may still lose to a well-structured human team.