The Economics of Intelligence: A DeepMind Institute Essay on Firms, Wages and Agents
In an experiment described in the essay, 100 people took a quiz covering math, verbal reasoning, science, patterns, culture and economics. Scores in one domain were a weak predictor of scores in another. Pairs chosen for complementary strengths beat randomly chosen pairs by 32% on a task, according to the authors. Alex Imas, Andrew Koh and Shane Legg of Google DeepMind build on that result in The Economics of Intelligence, published October 29 on the DeepMind Institute site. The formal model sits in a companion paper, Organizational Intelligence.
The authors say AI models are jagged in the same way, but models resemble one another, so adding copies of a model adds capacity without adding breadth. Individual benchmarks test isolated agents, while economic outcomes are set by competing organizations. A model can beat every human one-on-one and still lose to a well-structured team. The comparison they want measured is a team of AI agents against a team of humans or a mix of both.
Duplicability and coordination
The essay names two ways AI differs from human labor. Human expertise is in fixed supply, so wages rise as firms hire further up the talent distribution. AI can be replicated at the cost of compute, and where it matches human capability, the authors say human wages tie to that cost.
Human coordination is limited by attention, by tacit knowledge lost in handoffs, and by misaligned career incentives. AI agents can share identical weights and full working state, so those limits need not apply. The authors acknowledge that current agents coordinate poorly in practice, citing the CooperBench evaluation and a comment from OpenAI's Noam Brown that getting agents to coordinate productively is very difficult. They compare the potential to the telegraph and the railroad, which reorganized production through better handoffs.
Organizational scaling laws
The authors say competence sets output per worker at a fixed firm size, and coordination sets how output changes as the firm grows. They describe three phases. Below a coordination floor, solo work beats teams. In the middle phase, output per person rises with division of labor and then declines as the firm grows, and better coordination lets firms get bigger before that decline sets in. Past what the essay calls a supercooperative threshold, each added member raises productivity, and firm size is limited only by the size of the market.
On this account, competence moves the economy gradually and coordination moves it in jumps, with sudden shifts in concentration, employment and wages.
Four regimes
Under minimal AGI, AI matches average human capability across tasks. Market structure barely changes because balanced human teams still win in deep markets, but wages move first. Pay for AI-matched workers falls toward compute cost while specialists and coordinators earn more. The authors conclude that employment statistics are weak evidence that AI has had no effect.
Under supercompetent AI, models beat every human expert at everything except coordination. Firms become more productive and leaner but not larger. Pay for routine specialist work declines, and human value moves toward judgment and orchestration.
Under supercooperative AI, minimal-level competence combines with coordination past the threshold. In deep standardized markets such as software, logistics and finance, AI-only firms grow and abruptly outcompete smaller ones. Niche markets with bounded demand resist concentration, though workers moving into them push wages down.
Artificial superintelligence passes both thresholds. Lean AI firms can serve niche demand without scale, which removes that shelter for human teams. The authors suggest spending may flow toward a relational sector, including healthcare, education and craft, where human involvement is part of what buyers value, and toward new jobs that are hard to name today.
What to measure
The authors read current conditions as consistent with the framework: models improve quickly in competence, lag in coordination, and agent swarms show diminishing returns. They point to leaner startups, flatter org charts, softer entry-level hiring and steady or rising demand for experienced people who direct complex work.
They call for benchmarks that score organizations of agents on dividing unfamiliar problems without duplication, sharing context cleanly across handoffs, and catching errors over long horizons instead of compounding them. They also want economic statistics to track compute spending next to payroll. The full essay is at institute.deepmind.com/essays/the-economics-of-intelligence. Wages and job roles, in their view, are the earliest signals of change.
David Borish is the author of The Tony Hawk Paradox: When Video Games Predict Reality. He writes at davidborish.com
