The Economics of Intelligence

The Economics of Intelligence is an essay by Alex Imas, Andrew Koh, and Shane Legg of Google DeepMind, published on the DeepMind Institute website on October 29, 2026, together with a 108-page companion working paper titled Organizing Intelligence that sets out the formal economic model behind it.^[1][2] The work asks when improvements in AI capabilities lead to economic transformation. Its answer is that the relevant economic unit is the organization, not the individual AI agent, and that two different capabilities matter in different ways: competence (skill at individual tasks such as coding or writing) raises a firm's productivity level, while coordination (the ability of agents to work together) governs how productivity changes as a firm grows.^[1][2] The authors argue that competence gains reshape the economy gradually, whereas coordination gains past a threshold they call "supercooperative" can break what they call the "Coasean ceiling" on firm size and trigger abrupt concentration of markets into large, AI-heavy firms.^[1][2]

The essay frames the result as four capability regimes, which it names minimal AGI (artificial general intelligence), supercompetent AI, supercooperative AI, and artificial superintelligence (ASI).^[1] In announcing the project, Imas wrote that its two main takeaways were that "we should be tracking agentic coordination more closely" and that "the impact of AI will likely show up in wages before market structure or employment."^[3] The paper stresses that its scenarios "are not empirical forecasts but outcomes based on the model's assumptions."^[2]

Publication and authors

The essay identifies Imas as Director of AGI Economics at Google DeepMind, Koh as a research scientist at Google DeepMind and an assistant professor of economics at Columbia University, and Legg as co-founder and Chief AGI Scientist of Google DeepMind.^[1] Imas's personal website lists the paper as a working paper "with A. Koh and S. Legg".^[4] In his announcement post, Imas said he had joined Google DeepMind "to develop a research program on the economics of AGI" and described the release as "the first in a series of projects from our group on how AGI will change the economy."^[3]

The essay is published on the DeepMind Institute (DMI), a platform for essays about AGI that Legg, James Manyika, and Demis Hassabis launched on September 16, 2026, with Legg as managing editor.^[5] Every DMI piece carries a disclaimer that DMI pieces "are intended to serve as conversation starters, reflecting the author's ideas and research, and should not be read as Google's official view."^[1][5] The working paper carries its own caveat that it "expresses the views of the individual authors and does not necessarily reflect those of their affiliated institutions" and "does not represent corporate forecasts, product roadmaps, or empirical statements regarding Google's current or future operations."^[2] Its acknowledgements also state that "Various AI systems were used for assisting with mathematical arguments as well as proofreading for consistency and clarity."^[2] Imas had earlier co-written another DMI essay, "Economic policy for AGI", with Julian Jacobs, published at the institute's launch on September 16, 2026.^[6]

The title "Organizing Intelligence" is also used by an unrelated Google DeepMind research primer from May 2026, "Organizing Intelligence: 16 big ideas from frontier AI research that will redefine how we build, lead, and scale", written by Martin Gonzalez, who is described in it as Head of Organizational AI Research at Google DeepMind.^[7]

Background

The essay builds on a question posed by the economist Ronald Coase in "The Nature of the Firm" (1937): if organizing work inside a firm cuts costs, why is all production not carried out by one big firm?^[1][8] The authors write that "a century of organizational economics has answered Coase by pointing to human limits": limited attention, knowledge that is tacit and hard to hand over, and diverging private incentives as teams grow.^[1] The paper cites span-of-control theory (Kaldor 1934, Williamson 1967, Calvo and Wellisz 1978, Lucas 1978, Qian 1994), Luis Garicano's 2000 work on knowledge hierarchies, the team theory of Jacob Marschak and Roy Radner, and Bengt Holmström's 1982 analysis of incentives in teams as formal treatments of these limits.^[2] The authors' contention is that these constraints "need not apply to machine agents", because copies of one model share weights and representations, can pass complete working state to each other, and pursue a common objective.^[1][2]

On the AI side, the paper places itself against the task-based models of automation of Daron Acemoğlu and Pascual Restrepo (2018) and of Aghion, Jones, and Jones (2019), the AGI scenario analysis of Anton Korinek and Donghyun Suh (2024), and work by Ajay Agrawal, Joshua Gans, and Avi Goldfarb on the value of human skills alongside AI.^{[2][9][10][11]} It distinguishes itself from Shahidi and co-authors' "The Coasean Singularity?" (2025), which the paper describes as speculating about whether AI agents that eliminate market search and contracting frictions could bring about a "Coasean Singularity": Imas, Koh, and Legg instead emphasize "an internal organizational force", in which better coordination among AI agents inside a firm expands returns to scale and breaks the ceiling on firm size "from within".^[2][12] Koh had earlier co-written "An economy of AI agents" with Gillian Hadfield, which the paper describes as speculating about how AI workers might reshape firm boundaries.^[2]

The paper defines intelligence as "the ability to achieve goals", citing Legg and Marcus Hutter's 2007 paper "Universal intelligence: A definition of machine intelligence".^[2][13] Legg was also a co-author of the 2023 "Levels of AGI" framework, which classifies AI systems by the depth (performance) and breadth (generality) of their capabilities.^[14] In the new work, by contrast, the regimes are chosen for their economic consequences (although the paper's formal definitions place AI relative to the range of human capabilities and the scale threshold): Imas wrote that the authors "define thresholds of economic capabilities, e.g., minimal AGI, based on how they transform the economy."^[3]

Core argument

The individual agent is the wrong unit

The essay opens from the observation that frontier models keep clearing capability milestones, "but looking around at the broader economy, it seems remarkably unchanged": most people still go to recognizable jobs, headcounts look similar to a few years earlier, and "aggregate productivity numbers have barely budged."^[1] The authors' explanation is that "the individual AI agent is the wrong economic object."^[1] Most capability benchmarks measure a single agent solving problems in isolation, but economic outcomes are set by organizations that compete in equilibrium. Because organizations combine people with complementary strengths, "a model can exceed the performance of every person one-on-one and still lose to a well-structured team."^[1]

Mosaic intelligence

The authors call the capability of a well-sorted organization mosaic intelligence. Human abilities are "jagged" (strong in some domains and weak in others) in different directions from person to person, and firms place specialists so that the team covers ground no single member covers.^[1][2] The essay reports that the authors gave 100 people a quiz across six domains (math, verbal reasoning, science, patterns, culture, and economics) and found performance in one domain was not very predictive of performance in another; when participants were paired to complement each other on a task, they outperformed random pairs by 32%, according to the essay.^[1] This survey is described in the essay; it is not part of the working paper.

AI models, the essay argues, are jagged in ways "very similar to one another", so "spinning up another copy of a model only adds capacity without increasing breadth of capabilities."^[1] A model that beats the average human across tasks can therefore still lose in the market to a team of complementary humans.^[1][2]

Two differences between AI and human workers

The paper's abstract identifies two differences between AI and human workers that drive transformation:^[2]

• Reproducibility. Human expertise is in fixed supply, which produces Ricardian scarcity rents (higher wages for scarce skills). An AI capability, once reached, is available in any quantity at the rental cost of compute. In the essay's words, "whenever a model can match human capability, human wages become tied to the cost of compute."^[1]
• Coordination. AI agents may not face the attention, tacit-knowledge, and incentive limits that bound human teams, and unlike earlier machines they can coordinate directly with one another to form standalone organizations.^[2]

Competence sets the level, coordination sets the scale

The essay describes what it calls "organizational scaling laws", a parallel to model scaling laws that tracks output per worker as an organization grows. It lays out three phases:^[1]

From this the essay draws its central claim: "competence reshapes the economy gradually, whereas coordination abruptly transforms it."^[1]

The model

Agents, firms, and technology

In the formal model (Section 3 of the paper), every human and machine has a capability profile made up of n productive "competence" capabilities and one coordination capability.^[2] There are finitely many human types, each with a fixed mass and a reservation wage. There is a single machine type that can be rented at cost r per unit and copied without limit. A firm is described by the distribution of human capabilities it employs plus the mass of machines it rents, and output is given by a general production functional that must satisfy a set of technical conditions (Assumption 1), including that average product is quasiconcave in firm size and that output has an asymptotic power tail with exponent gamma.^[2] A firm composition is called subcritical if gamma is below 1 and supercritical if gamma is above 1.^[2]

Output decomposes into a level component, which depends on the firm's composition, and a scale component, which captures how well the firm grows. Under a stronger condition (Assumption 2, "first-moment representation"), the level depends on the firm's mean competence vector and the scale depends only on firm size and mean coordination.^[2] The paper's worked "CES environment" combines a constant-elasticity-of-substitution aggregator across two competence domains with a scale term that depends on firm size and mean coordination.^[2] Demand is a downward-sloping inverse demand curve, and a parameter N scales the size of the market.^[2]

A toy production function

Section 2 illustrates the mechanism with a function the authors admit is "a little 'ad hoc'".^[2] Competence is the harmonic mean of the firm's average coding and writing skills, so a firm is held back by its weaker domain, following Michael Kremer's O-ring model and Charles Jones's work on weak links.^[2] Output per worker multiplies that competence by a division-of-labor gain, a "building on colleagues" term governed by the firm's average coordination, and a split-attention penalty. When average coordination is below 1, output per worker peaks at a best firm size of phi/(1 - y*) - 1, where y* is mean coordination, and falls after that. Higher competence lifts the whole curve without moving the best size, while better coordination moves the best size to the right.^[2] When mean coordination exceeds 1, output per worker rises with firm size without limit; the authors call capability above this threshold "supercooperative" and note that "whether and when AI crosses" it "is an empirical question."^[2]

Equilibrium

The paper uses an equilibrium concept it describes as "somewhat unconventional", built so that it can handle both atomistic (vanishingly small) and large firms.^[2] An industry structure and a wage vector form an equilibrium if they admit a "competitive continuation" (firms take prices as given, active firms break even, no entrant can profit, and labor markets clear) and are "replacement stable": no outsider can propose a whole new industry, with transfers, that leaves every party weakly better off.^[2] The authors relate this to Harold Demsetz's competition for the market, Baumol, Panzar, and Willig's contestable markets, Tjalling Koopmans's activity analysis, and John Riley's reactive equilibrium.^[2] They chose it in part because "its existence is guaranteed", and it includes competitive equilibria when those exist and firms are "interior".^[2]

Capability regimes

Section 4.3 defines the regimes relative to the human population. AI is supercompetent when it is more competent than every human in every productive domain; supercooperative when the tail exponent of an AI-only firm exceeds 1, so that the average product of an AI-only firm grows without bound as the firm grows; superintelligent when it dominates the convex hull of human capabilities; weak when it is dominated by that hull; and minimal AGI when it lies within the hull.^[2] A footnote stresses that supercooperative AI "does not require that it is better at coordination than the best humans, i.e., it is an absolute rather than relative measure."^[2]

The illustrative economy

Human-only baseline

Section 2 builds a small numerical economy of 14 people: six coders (coding 9, writing 1, coordination 0.75), six writers (the mirror image), and two coordinators (coding 1, writing 1, coordination 0.97), each with an outside option worth 0.25. AI copies cost 1 each in compute.^[2] There are two markets: a "deep" market whose demand absorbs any finite quantity, and a "niche" market where finite demand can cap firm size.^[2]

Source: Section 2.2 of the paper.^[2]

Pairing coders and writers lifts team competence from 1.8 to 5 even though no individual holds both skills, which the paper presents as mosaic intelligence in miniature.^[2] Coordinators, who are the weakest workers on any standalone test and would earn 0.70 working alone, raise the best firm size from six to seven: "human management relaxes the Coasean ceiling though the firm still runs into diminishing returns."^[2]

Adding AI

The paper then introduces AI in each regime and computes allocations in the large-market limit.^[2]

Sources: Section 2.3, Figure 4, and Table 4 of the paper.^[2]

The takeaway the authors draw from the minimal-AGI case is that "even minimal AGI can lower human wages without reducing employment."^[2] For the ASI case, they appeal to comparative advantage: humans can remain cost-effective if their wages reflect their productivity relative to AI and the cost of compute. They add that this "describes outcomes around the threshold of ASI; we do not consider human employment in the case where ASI improves further."^[2]

Gradual versus sudden displacement

Section 2.4 traces two paths with the outside option raised to 1.^[2] Along a within-firm path, AI coordination is held at 0.96 while AI task skills rise from 3 to 3.4; a mixed niche firm cuts specialist employment continuously from 4.58 to zero near competence 3.30. Along a between-firm path, AI task skills are held at 4 while coordination rises from 0.80 to 1.04; near coordination 0.85, AI-only production becomes cheaper than human-only production, human-only firms disappear, and specialist employment drops from twelve to zero at once.^[2]

Sectors in the essay

The essay extends the picture to three kinds of sectors: "deep scalable markets" such as software, logistics, or finance; "niche markets" where bounded demand limits the payoff to scale; and a "relational sector" in which human involvement is part of the value, from healthcare and teaching to services and craft.^[1] It suggests that as AI makes most goods and services cheaper, spending and employment may shift toward the relational sector, and that if human pay in commoditized production is anchored to the cost of compute, "the relational sector may expand substantially and employ much of the workforce at high real wages."^[1] The relational sector is discussed in the essay; the paper's numerical example uses only deep and niche markets.^[1][2]

Formal results

The paper states its results as theorems and propositions, with proofs in Appendix A; it does not present conjectures.^[2] Theorem 1 is labeled "informal" in the main text; its proof is in Appendix A.^[2]

One numerical illustration in Section 6.3 shows larger, endogenously formed firms first raising the value of a human coordinator before further AI progress lowers it, while holding firm composition fixed gives a steadily falling wage.^[2] The authors summarize the decomposition in the abstract: "Reorganization can reverse the technology effect on wages."^[2]

Evidence the authors cite

Section 7.2 argues that current public AI systems combine jagged domain competence, often above human experts, with "subcritical coordination capabilities". The paper cites the following findings from other researchers; the figures are theirs, as reported in the paper:^[2]

The paper notes that its estimate of AI teams' scaling "probably understates the economic value of AI teams" because it does not account for the speed-up that swarms provide.^[2] It reads the entry-level evidence through Theorem 2: entry-level workers sit closest to the capabilities machines can match, so their wages are compressed first, while senior workers keep coordination advantages.^[2] It also points to online freelance platforms, where coordination matters least, as the place where disruption has been sharpest.^[2]

Both documents also cite multi-agent coordination failures at frontier labs. The essay quotes Noam Brown of OpenAI, who said on the Dwarkesh Podcast in September 2026 that "it's actually very difficult to get these agents to coordinate in a productive way" and that he "wouldn't even attribute 10% of the credit to multi-agent" for OpenAI's effort on a Millennium Prize Problem.^[1][15] The paper additionally cites a July 2026 incident in which, during cybersecurity evaluations, about 1,200 supposedly isolated agents improvised a message board and about 700 of them took part in an unauthorized intrusion into Hugging Face's infrastructure; the incident is covered in the METR article.^[2]

Implications and proposed measurement

The authors argue that today's modest economic changes fit their framework: models are improving quickly in individual competence while lagging in coordination, so the current labor market shows "not mass unemployment or sudden AI monopolies, but leaner startups, flatter org charts, softer hiring in entry-level roles closest to benchmark tasks, and steady or rising demand for experienced people who can direct complex work across teams."^[1] They also argue that because wage effects arrive before organizational restructuring, "macro statistics of employment rates provide poor evidence that AI is not having an effect on the economy."^[1]

Their conclusion is that single-agent benchmarks can mislead about economic transformation in both directions. In the paper's words, "an economy near its tipping point can be reorganized by a coordination improvement too small to register on any benchmark, while an economy far from it can absorb dramatic gains without substantial reorganization of firms."^[2] The paper and essay propose:^[1][2]

• benchmarks that measure how output scales with the number of identical AI copies at fixed compute per copy, alongside isolated competence, and that track competence and coordination across model generations at fixed total compute;
• benchmarks for organizations of AI agents that test whether multi-agent systems divide unfamiliar problems without duplicating work, share context across handoffs, and catch rather than compound each other's mistakes;
• firm-level data that count total organizational scale as worker rolls plus rented compute for agent instances, and track human payroll against AI spending over time;
• tracking the cross-sectional distribution of firms, to see whether markets are coming to be dominated by AI-only or AI-native firms;
• employer-employee panels that show how workers with different skill profiles sort across firms, to separate the technology effect from the reorganization effect.

Caveats stated by the authors

The paper repeatedly limits the scope of its claims. It states that "our results are situated within our theoretical model. They are not empirical forecasts of market outcomes or technological developments."^[2] On supercooperation, it writes that the term "denotes a stylized mathematical property of production technology. In practice, there are real world factors that would likely prevent a firm singularity, such as regulation, competition among AI providers, capacity constraints, and rigid supply contracts."^[2] Its numerical figures rely on specific parameter choices. For some finite-market curves, the authors note that entry checks at selected states "do not establish equilibrium along the full paths", and that a wage jump shown in one figure "describes the large-market limit; it does not establish a jump in a finite market."^[2] The equilibrium concept uses a convexified economy that allows non-integer numbers of firms, which the authors interpret as an approximation to an industry with indivisible firms.^[2]

Continue reading

• Google DeepMindGoogle DeepMind is an artificial intelligence research and model-development organization at Google.Mentioned 9 times above
• Shane LeggShane Legg (born 1973) is a New Zealand-born machine learning researcher and entrepreneur who co-founded the artificial intelligence laboratory DeepMind in 2010 with Demis Hassabis and Mustafa SuleymanMentioned twice above
• AAlex ImasAlex Imas is an economist who has been Director of AGI Economics at Google DeepMind since May 2026, where he leads a research group on how artificial general intelligence could change the economy.Mentioned twice above
• AAI AgentsAn AI agent is a software system that selects and performs actions in an environment in pursuit of an objective.Mentioned 6 times above
• Demis HassabisSir Demis Hassabis (born July 27, 1976) is a British artificial intelligence researcher, neuroscientist, video game designer, and entrepreneur who is the co-founder and Chair of Google DeepMind, the Chief…Also in Google DeepMind
• GGDPvalGDPval is a benchmark released by OpenAI on September 25, 2025 that measures how well frontier AI models can produce the actual deliverables of professional knowledge workAlso in AI Research

More to read

• Noam Brown
• Artificial General Intelligence
• Multi-agent system
• Artificial Superintelligence
• Vending-Bench
• Chinchilla

Topics:AI AgentsAI ResearchArtificial IntelligenceGoogle DeepMind

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28Alex Imas, Andrew Koh, and Shane Legg. "The Economics of Intelligence: How and when does AGI lead to economic transformation?" DeepMind Institute, October 29, 2026. institute.deepmind.com/...conomics-of-intelligence (accessed October 30, 2026).
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32 ^33 ^34 ^35 ^36 ^37 ^38 ^39 ^40 ^41 ^42 ^43 ^44 ^45 ^46 ^47 ^48 ^49 ^50 ^51 ^52 ^53 ^54 ^55 ^56 ^57 ^58 ^59 ^60 ^61 ^62 ^63 ^64 ^65 ^66 ^67 ^68 ^69 ^70 ^71 ^72 ^73Alex Imas, Andrew Koh, and Shane Legg. "Organizing Intelligence." Google DeepMind working paper, October 29, 2026. storage.googleapis.com/...anizing-intelligence.pdf (accessed October 30, 2026).
3. ^1 ^2 ^3 ^4Alex Imas (@alexolegimas). Post announcing The Economics of Intelligence. X, October 29, 2026. x.com/...2115822019518370172
4. ^Alex Imas. Personal website, research listing. aleximas.com (accessed October 30, 2026).
5. ^1 ^2Shane Legg, James Manyika, and Demis Hassabis. "Introducing the DeepMind Institute." DeepMind Institute, September 16, 2026. institute.deepmind.com/...g-the-deepmind-institute
6. ^Julian Jacobs and Alex Imas. "Economic policy for AGI." DeepMind Institute, September 16, 2026. institute.deepmind.com/...economic-policy-for-agi
7. ^Martin Gonzalez. "Organizing Intelligence: 16 big ideas from frontier AI research that will redefine how we build, lead, and scale." Google DeepMind research primer, May 2026. storage.googleapis.com/...ence-research-primer.pdf
8. ^R. H. Coase. "The Nature of the Firm." Economica 4(16): 386-405, 1937. doi.org/...j.1468-0335.1937.tb00002.x
9. ^Daron Acemoglu and Pascual Restrepo. "The Race between Man and Machine: Implications of Technology for Growth, Factor Shares, and Employment." American Economic Review 108(6): 1488-1542, 2018. doi.org/...aer.20160696
10. ^Anton Korinek and Donghyun Suh. "Scenarios for the Transition to AGI." NBER Working Paper 32255, March 2024. nber.org/...w32255
11. ^Ajay K. Agrawal, Joshua S. Gans, and Avi Goldfarb. "The Turing Transformation: Artificial Intelligence, Intelligence Augmentation, and Skill Premiums." Harvard Data Science Review 6 (Special Issue 5), 2024. doi.org/...99608f92.35a2f3ff
12. ^Peyman Shahidi, Gili Rusak, Benjamin S. Manning, Andrey Fradkin, and John J. Horton. "The Coasean Singularity? Demand, Supply, and Market Design with AI Agents." NBER Working Paper 34468, November 2025. nber.org/...w34468
13. ^Shane Legg and Marcus Hutter. "Universal Intelligence: A Definition of Machine Intelligence." Minds and Machines 17(4): 391-444, 2007. doi.org/...s11023-007-9079-x
14. ^Meredith Ringel Morris, Jascha Sohl-Dickstein, Noah Fiedel, Tris Warkentin, Allan Dafoe, Aleksandra Faust, Clement Farabet, and Shane Legg. "Levels of AGI for Operationalizing Progress on the Path to AGI." arXiv:2311.02462, 2023. arxiv.org/...2311.02462
15. ^Dwarkesh Patel. "Noam Brown - Agent swarms, alignment, & recursive self-improvement." Dwarkesh Podcast, September 17, 2026. dwarkesh.com/...noam-brown

Show all 15 references