The Pulse
DeepMind Authors Say Coordination Could Reshape AI’s Economic Impact
An essay from the DeepMind Institute argues that AI’s economic effects depend not only on individual capabilities but also on whether agents can coordinate within organizations. It lays out possible changes to firms, wages and employment ac
AI.info Team · Written with AI · fact-checked ·
AI systems keep clearing individual skill tests, but the wider economy has yet to show the sweeping changes those scores might suggest. In an essay published October 29, Google DeepMind researchers Alex Imas, Andrew Koh and Shane Legg argue that the gap reflects a mismatch: benchmarks measure agents working alone, while economic competition plays out among organizations.
The researchers’ claim is not simply that more capable models will raise productivity. They argue that agents’ ability to divide work, share context and coordinate could determine whether AI produces gradual shifts or a sudden reorganization of firms, jobs and wages. Their essay presents scenarios derived from an economic model, not a finding that such a transformation has already occurred.
Why solo benchmarks miss the team
People bring different strengths to a workplace. A company can combine specialists, assign tasks to those best suited to them and use managers to keep handoffs on track. A model might outperform any one employee on a test and still fall short of a team whose members complement one another.
To illustrate that point, the authors describe a quiz given to 100 people across six domains: math, verbal reasoning, science, patterns, culture and economics. They report that participants paired to complement one another performed 32% better than random pairs on a task. That result illustrates their argument about complementary skills; it does not measure AI’s effect on wages, employment or productivity across the economy.
The authors say AI copies can be supplied as compute allows, unlike human workers, whose time and availability are limited. But duplicating one model adds capacity without necessarily adding new strengths. Their economic comparison is therefore between organizations of people and organizations of AI agents—not between one model and one person.
Coordination sets the size of the firm
In the framework, competence raises output per worker at a given firm size. Coordination affects what happens as a firm adds more workers or agents. Weak coordination makes collaboration inefficient; better coordination can help a company grow before communication friction and misaligned incentives weigh on output.
The authors describe a threshold beyond which each additional agent makes the organization more productive than the last. They call that point “supercooperative.” If an AI firm crosses it in a large, standardized market, the model suggests that it could grow rapidly and outcompete smaller rivals, even if its individual agents are not extraordinary on isolated tests.
“Put another way, competence sets the potential productivity level; coordination determines how big an organization can get without becoming too messy to function.”
The authors: Alex Imas, Director of AGI Economics at Google DeepMind; Andrew Koh, Research Scientist at Google DeepMind; and Shane Legg, Co-founder and Chief AGI Scientist at Google DeepMind.
The authors contrast that possibility with a world where models become much more capable but coordination stays limited. In that scenario, companies could produce more with smaller teams, while industries remain divided among many firms. Routine specialist wages might fall toward the cost of compute, while human workers who direct and align complex projects retain an advantage.
Four regimes, different labor outcomes
The essay maps possible outcomes across four regimes: minimal AGI, supercompetent AI, supercooperative AI and artificial superintelligence. The distinctions turn on whether AI matches or exceeds human skills and whether agents can coordinate beyond the proposed threshold.
For minimal AGI, the authors expect wage changes to precede major shifts in firm structure: workers whose skills AI can reproduce face pressure, while specialists and coordinators may command more. Supercompetent AI, in their account, could make firms more productive without making them larger. Supercooperative systems could instead drive concentration in scalable markets; the combined supercompetent and supercooperative regime extends that pressure to smaller, specialized markets.
Those are conditional projections, not a calendar or a forecast that any particular threshold is near. The essay itself says current AI systems coordinate poorly in practice. It also proposes new measurements: tests of how agents divide unfamiliar work, pass context between tasks and catch one another’s errors, alongside economic tracking of payroll and compute spending.
The essay is a framework, not a Google position
The essay says it draws on a separate paper, “Organizing Intelligence,” which presents the full economic model. A disclaimer on the DeepMind Institute page says its essays are intended as conversation starters reflecting their authors’ ideas and research, and should not be read as Google’s official view.
That distinction matters because the essay’s central contribution is a way to frame the question, not evidence that employment or market concentration has already shifted because of agent coordination. Its proposed early indicators are wages and changing job roles; its most dramatic outcomes depend on AI systems crossing a coordination threshold the authors say today’s systems have not reached.