An essay from the Google DeepMind Institute proposes that the economic impact of advanced artificial intelligence will depend not only on how capable individual agents become, but on how effectively they coordinate with one another. The authors frame economic transformation around capability thresholds that affect the structure of firms, employment and wages, drawing on a formal model presented in a paper titled “Organizing Intelligence.”
The central argument is that familiar AI benchmarks measure the wrong unit for forecasting economy-wide change. Most tests evaluate a model working alone on a defined problem. Economies, by contrast, organize people into firms whose members specialize, exchange work and combine complementary abilities. A system that outperforms each individual worker on isolated tasks could still be less productive than a well-structured human team.
The essay calls this collective advantage “mosaic intelligence.” To illustrate it, the researchers gave 100 people a quiz spanning mathematics, verbal reasoning, science, pattern recognition, culture and economics. Participants showed uneven strengths across subjects. When people were paired to complement one another on a task, the essay reports that they outperformed random pairs by 32 percent. The result is used as an example of how organizations can cover individual weaknesses by matching specialists to the work they do relatively well.
Coordination also creates value for roles that may look unremarkable on standalone benchmarks. A person with average coding and writing ability, for example, could improve the output of specialists by managing handoffs and reducing friction. The authors argue that tests focused on solo performance generally fail to capture that contribution.
Machine intelligence could eventually alter this organizational equation in two ways. AI agents can be replicated as long as additional compute is available, unlike the relatively fixed supply of human workers. Machine teams may also face different constraints on communication, incentives and management span. If agents can collaborate directly instead of operating individually inside human organizations, the essay contends, they would become a new kind of economic input rather than simply another productivity tool.
The framework describes regimes including minimal AGI, supercompetent AI, supercooperative AI and artificial superintelligence. These labels are analytical categories, not announcements that current systems have reached them. Their purpose is to distinguish improvements in individual competence from improvements in scalable cooperation.
The analysis offers a possible explanation for why rapid gains on frontier benchmarks have not yet produced equally dramatic changes in aggregate productivity or employment. Its conclusion is a model-based forecast rather than an observed economic outcome: visible disruption may accelerate only after systems cross coordination thresholds that today's benchmarks rarely measure.
