However Strong AI Is at Individual Tasks, It’s Wasted If Organizations Can’t Coordinate

Over the past few years, models have topped benchmarks every few months: writing production code, solving math problems, running scientific research workflows. But look up at the macro data, and employment, firm boundaries, and productivity have shown no corresponding dramatic changes. On October 29, the DeepMind Institute released 'The Economics of Intelligence,' in which Alex Imas, Andrew Koh, and Shane Legg offer an explanation: comparing a single AI with a single person chooses the wrong unit for economic analysis. The unit of competition among firms is the organization.

The paper breaks the problem into two capability thresholds: competence—individual skills such as programming, mathematics, and writing; and coordination—the ability of agents to divide labor, hand off work, share context, and correct one another. The former determines the output level of each person in an organization; the latter determines how large an organization can grow without falling apart.

One detail deserves technical readers' attention: the paper had 100 people take tests in six domains and found that people's abilities across domains are distributed in a very 'jagged' way—someone good at mathematics is not necessarily good at pattern recognition. The role of the organization is to assemble complementary experts into a mosaic of capabilities that covers one another's gaps. In the experiment, people paired by complementarity performed 32% better than those paired randomly. A single model can beat every individual, yet it may lose to a complementary human team.

The paper uses an organizational scale curve to describe three stages. When coordination is below a baseline, a single agent is more cost-effective at doing the work. When coordination is in the middle range, division of labor can raise output per person, but as the organization grows, communication friction increases; the curve flattens and then declines—this is the world of most companies today. Once coordination passes a critical point, each additional agent makes the team more efficient instead; returns to scale shift from diminishing to increasing, the Coasean ceiling is broken, and firms can grow until they are limited only by market size.

Four Regimes

Depending on whether the two thresholds have been crossed, the paper delineates four economic regimes: minimal AGI, supercompetent AI, supercooperative AI, and ASI.

Minimal AGI refers to replicable AI reaching average human capability. On the surface, the economic structure barely changes, and complementary human teams can still beat AI firms with average capability. Change appears first in wages: for skills that AI can match, wages are pressed down toward computing costs; scarce experts and coordinators earn higher wages. Small AI firms enter price-sensitive and niche markets, pushing down the prices of goods.

Supercompetent AI refers to AI that surpasses all human experts in individual capabilities but still coordinates poorly. The result is somewhat counterintuitive: firms become leaner and more efficient, employee counts may fall, and industries remain fragmented. The paper calls it 'abundance without concentration.' The human role shifts toward judgment, orchestration, and setting direction for teams.

Supercooperative AI refers to AI whose individual capabilities are only at minimal AGI, but whose agent coordination has crossed the threshold. Deeply standardized markets will suddenly reorganize; a few large AI organizations rapidly capture share, and wages fall sharply. Niche markets are protected by demand ceilings, but labor displaced from other sectors will push down wages in those markets and in relational service industries.

ASI refers to passing both thresholds. Deep markets are dominated by large AI companies, and niche markets can no longer protect human teams through small scale; lean AI firms can serve specialized needs. The paper argues that employment may flow toward relational sectors such as healthcare, education, and services, where human involvement itself constitutes value. Historical experience says new jobs appear after automation, but this time the scope of automation may be broader.

Wages Move Before GDP

The paper's core judgment is that improvements in competence smoothly change prices, wages, and the human–AI mix; a breakthrough in coordination suddenly changes market structure, employment, and wages. Therefore, wage effects will precede organizational restructuring. The fact that the macro employment rate has not moved cannot prove that AI has had no impact.

Benchmarks Miss Coordination

Most existing benchmarks measure a single agent's performance in isolated environments. The paper argues that this overestimates models' competitiveness against complementary human teams and also misses early warnings of approaching the coordination threshold. What is needed are organization-level benchmarks: can multiple agents decompose unfamiliar problems without duplicating work, can they hand off context cleanly, and can they discover and correct one another's errors over long horizons?

In an interview, Noam Brown mentioned that when thousands of agents work in parallel on open mathematics problems, it is very hard to get them to coordinate effectively, with multi-agent contributions accounting for less than 10%. This is consistent with the paper's judgment: individual capabilities are rising, while coordination capabilities remain slow.

History offers a reference point as well. In the 19th century, the telegraph and railways did not make individual workers stronger; they increased the speed and fidelity of handoffs, and modern corporate structures were reorganized as a result. What to watch today is likely the organizational forms of multi-agent coordination and early changes in wage structures; single-agent SOTA is no longer enough.

Full paper: https://institute.deepmind.com/essays/the-economics-of-intelligence/