betaFeedback

WorksEconomics, Econometrics and Finance

Organizing Intelligence

Abstract

When will advances in AI lead to economic transformation? We develop a model of how intelligence, the ability to achieve goals, matters for organizations, the key unit of economic production. Agents (human or AI) with multidimensional capabilities sort endogenously into competing firms. A firm's productivity depends on its agents' capabilities; we distinguish competence, which shifts productivity levels, from coordination, which governs how productivity scales with size. Two differences between AI and humans drive transformation: AI expertise is reproducible while humans' are in fixed supply, and AIs may not face the same human limits on coordination. We analyze how equilibrium market structure, productivity, employment, and wages evolve as AI capabilities improve. When AI is more capable than the average human, organizations of complementary humans can nonetheless outperform AI. Increases in AI competence can change the composition of workers within firms and decrease firm size. As AI coordination improves, however, this can break the Coasean ceiling on firm size and the market structure undergoes a phase transition toward large, concentrated AI-heavy firms. Changes in human employment are gradual when AI displaces workers within mixed human-AI firms, or sudden when AI firms outcompete human firms. We decompose wage changes into a technology effect, which changes the value of humans within existing firms, and a reorganization effect as firms adjust size and composition. Reorganization can reverse the technology effect on wages.

We thank Allan Dafoe, Matt Elliott, Joshua Gans, Gillian Hadfield, Jacopo Perego, Andrea Prat, Eduard Talamas, Phil Trammell, and Birger Wernerfelt for helpful conversations and suggestions. We also thank Pedro Aldighieri for superb research assistance. Various AI systems were used for assisting with mathematical arguments as well as proofreading for consistency and clarity.

The paper expresses the views of the individual authors and does not necessarily reflect those of their affiliated institutions. This research is intended solely for academic discussion and does not represent corporate forecasts, product roadmaps, or empirical statements regarding Google’s current or future operations.

Introduction

When will AI transform the economy? What kinds of firms will emerge, and what will happen to the returns to human expertise and to wages? We develop a model of how the capabilities of human and artificial intelligence interact with economic forces to jointly determine the composition and size of firms, surplus, and human wages. We model intelligence as the ‘ability to achieve goals’ [65]. This has typically been measured through isolated benchmarks that evaluate a single agent in isolation. But the economically relevant unit is the organization that competes in equilibrium under the forces of supply and demand [86].

Economic models often treat new technology either as a tool that makes workers more productive or as a factor that substitutes for humans task-by-task. But AI is different in three conceptually important ways. First, human capabilities are fixed, which can generate Ricardian scarcity rents in the form of higher wages for increasing expertise; by contrast, AI capabilities can be copied at the cost of compute which eliminates these rents. Second, human capabilities differ in jaggedness across people while copies of AI models overlap in their jaggedness; sorting heterogeneous workers into organizations creates what we call mosaic intelligence, such that human teams can out-compete more competent AI. Finally, unlike machines of the past that had to operate within human firms, AI agents can coordinate directly with one another to form standalone organizations. This can lead to rich non-monotonicities in market structure as AI capabilities increase.

We build a simple theoretical model incorporating those three features to formally characterize potential economic scenarios.2 For instance, as AI models become more competent at specific tasks, firms might first shrink as human managers can do more with AI agents. But past a coordination threshold, AI can break the Coasean ceiling and trigger a sudden phase transition toward a giant AI firm. Beyond this ceiling, firm sizes start to shrink continuously once again as fewer AI agents suffice to produce enough output to saturate the market.

In our model, each human and machine is endowed with a multidimensional capability profile. Multidimensionality captures an agent’s ability to achieve goals across different domains.3 Such agents—both human and machine—sort into firms whose output depends on both size, or how many workers the firm employs, and *composition of the capabilities of its constituent members. A central distinction is between competence, which captures domain-specific capabilities such as robotics, writing, or logical reasoning, and coordination, which captures capabilities that allow agents to work better with each other. This might comprise a better understanding of norms, reduced communication costs, reduced agency frictions, and so on. A firm with more competent agents enjoys a level shift in output. A firm with more cooperative agents, on the other hand, enjoys better returns to scaling—it can function better as the organization grows in size. In equilibrium, human and AI agents endogenously sort into competitive firms that serve a downstream market.4

How does the equilibrium market structure change as AI systems improve? Where are humans employed and how much do they get paid? This hinges on the direction of improvement, and whether specific capability thresholds are crossed. Our framework helps rationalize the why the rapid increases in AI capabilities across a wide array of domains can be accompanied by relatively little economic transformation. While current AI models are beginning to match or exceed human performance across many domains, they are still fairly poor at coordinating. Indeed, today’s labor market reflects such capabilities: no big effects on employment or market structure, but leaner startups, flatter org charts, and steady or rising demand for experienced managers who can direct complex work across teams.5 By taking seriously the endogenous organization of firms at different scales, our model makes nuanced and rich predictions on how different directions of AI improvements will shape surplus, employment, and wages and, in so doing, clarifies what it would take for economic transformation.

An illustrative example

Before turning to the general model, we build up a simple and concrete economy step-by-step to illustrate the key mechanisms of the paper.

The economy

Each worker (human or AI) brings capability vector y=(y1,y2,y∗)\mathbf{y}=(y_1,y_2,y_*), combining two productive task skills—coding y1y_1 and writing y2y_2—with a coordination capability y∗y_*. Let H\mathcal{H} denote the human population of NH=14N_H=14 workers across three roles (Table 1). Coders and writers are mirror-image specialists: each excels in one domain and struggles in the other. Coordinators are weak at both tasks but excel at facilitating cooperation across people. Any person who stays out of production takes an outside option worth w=1/4w=1/4. AI enters in § 2.3: firms can run any number of copies of an AI agent with profile yM=(y1M,y2M,y∗M)\mathbf{y}^M=(y_1^M,y_2^M,y_*^M) at a compute cost of r=1r=1 per copy.

RoleTask skillsCoordinationSupply
Coder9, 13/46
Writer1, 93/46
Coordinator1, 10.972

Table 1. Capabilities of agents in the economy. Task skills report coding y1y_1 first and writing y2y_2.

Competence. Competence governs how much output a firm produces per worker at a given size. A firm’s competence LL is the harmonic mean of its members’ average skills (yˉ1,yˉ2)(\bar{y}_1,\bar{y}_2) ([63, 56]):6

L=21/yˉ1+1/yˉ2.L=\frac{2}{1/\bar{y}_1+1/\bar{y}_2}.

Figure 1(a) plots this aggregator in skill space. Each curved contour traces skill pairs that deliver the same competence. Because the harmonic mean penalizes bottlenecks, a firm is held back by its weaker domain: solo coders and solo writers sit near opposite axes on a low contour (9/5=1.8)(9/5=1.8) despite excelling at one task, while a coordinator working alone sits on the lowest contour (1)(1). Scored in isolation without complementary sorting across people, the human population averages standalone competence and coordination

Figure 1. The capability envelope, competence, and firm scale.

LˉH=1NH∑i∈HL(y1i,y2i)≈1.69andyˉ∗H=1NH∑i∈Hy∗i≈0.78.\bar{L}_H=\frac{1}{N_H}\sum_{i\in\mathcal{H}}L(y_1^i,y_2^i)\approx1.69 \quad\text{and}\quad \bar{y}_*^H=\frac{1}{N_H}\sum_{i\in\mathcal{H}}y_*^i\approx0.78.

Coordination and firm size. Coordination governs how productivity changes as a firm adds workers. As Coase famously asked, why doesn’t one giant firm employ everyone? Larger firms divide work more finely, lifting output per worker, yet each person must also coordinate across more colleagues. Bounded human cognition and scarce attention mean that information degrades across handoffs and key details go unheeded ([90, 91, 69]); incentives across individuals also diverge as group size grows ([100, 47, 15]).

Consider a team building and documenting a software product. Two people working side by side can hold the full design in mind and resolve mismatches immediately. But expand that team to twenty people split across subteams, and members spend more time in meetings, meaning gets lost in handoffs, and subteams drift toward individual goals rather than the team’s mission, i.e., misaligned incentives. Classic span-of-control theory formalizes how finite managerial attention bounds firm size through these limits to coordination ([58, 100, 23, 67] (Qian 1994)). Garicano (2000) shows how hierarchies solve this issue by economizing on scarce attention by routing tasks within the firm.

Output per worker F/mF/m in a firm of mm workers combines these forces:

Fm=L⏟Firm comp.×eϕm/(1+m)⏟dividing work×(1+m)yˉ∗⏟building on colleagues×11+m⏟split attentionwhere ϕ=74.\frac{F}{m} = \underbrace{L}_{\text{Firm comp.}} \times \underbrace{e^{\phi m/(1+m)}}_{\text{dividing work}} \times \underbrace{(1+m)^{\bar{y}_*}}_{\text{building on colleagues}} \times \underbrace{\frac{1}{1+m}}_{\text{split attention}} \qquad\text{where } \phi=\frac{7}{4}.

Dividing work raises productivity quickly at small scale and then levels off. Meanwhile, each worker builds on colleagues’ output at rate (1+m)yˉ∗(1 + m)^{\bar{y}_*}, governed by the firm’s average coordination yˉ∗\bar{y}_*, while attention is strained in proportion to 1/(1+m)1/(1 + m). Of course this specification is a little ‘ad hoc’ but it will make the numbers concrete; our main model will impose more general assumptions on how the capabilities of individual workers matter for firm scaling.

Whenever yˉ∗<1\bar{y}_* < 1, bounded attention and potential misalignment eventually outweigh what each additional worker contributes to colleagues. At small scale, the steep gains from dividing work dominate that net coordination friction, so output per worker rises initially. Once division-of-labor gains level off, coordination friction takes over. Given these forces, output per worker peaks at the firm’s best size

m∗(yˉ∗)=ϕ1−yˉ∗−1,m^*(\bar{y}_*) = \frac{\phi}{1-\bar{y}_*} - 1,

and declines thereafter (Figure 1(b)). The key takeaway here is that higher competence shifts the level of the curve up while leaving best size unchanged, whereas better coordination increases returns to scaling, shifting best size to the right—the optimal size of the firm increases with coordination.

When yˉ∗>1\bar{y}_* > 1, the gains from building on colleagues’ work exceed the losses from splitting attention. Output per worker rises continuously with firm size rather than peaking and turning down (the dashed curve in Figure 1(b)); a larger firm always produces more per worker than a smaller one. For instance, Radner [83] shows that organizing information processing efficiently within the firm can generate increasing returns to scale (see also Bolton and Dewatripont [18]). There might also be informational economies of scale in which knowledge and ideas can be shared across the organization Wilson [101]. We call capability above this threshold supercooperative.

Human organizations run into decreasing returns for at least three reasons. First, human attention and memory are finite, creating constraints on the ability of agents to coordinate [91, 20]. Second, much of human knowledge is tacit [80], passing work from one person to another compresses rich internal context into lossy summaries [42]. Third, individual humans have private career concerns with their own objective functions, so incentive misalignment can compound with scale [47].

Reproducible AI agents need not share these constraints and frictions. Copies of a single model share identical weights and representations, can pass complete working state—entire codebases, intermediate derivations, and full context windows—without lossy translation, and execute a common objective. In large-scale software engineering or automated scientific discovery, thousands of parallel instances can explore distinct branches simultaneously while sharing verified state in real time. Whether and when AI crosses yˉ∗=1\bar{y}_* = 1 is an empirical question (see Ord, [74], for some early experiments).

Markets and equilibrium. Firms sell output into competitive markets with inverse demand

P(Q)=0.3(1−Q200N),P(Q) = 0.3\left(1 - \frac{Q}{200N}\right),

where QQ is total market output and NN measures market size.7 Our economy comprises two markets: a deep market (N→∞N \to\infty), where Q/(200N)→0Q/(200N) \to0 so buyers absorb any finite quantity at price 0.3, and a niche market (N=1N = 1), where buyer willingness to pay falls to zero at Q=200Q = 200 so finite market size can cap how large a firm grows. In our equilibrium concept (to be specified later) active firms break even, no entrant can earn positive profit, and competition selects the industry structure (collection of firms) that maximizes total surplus.8 Each worker’s equilibrium wage equals the shadow value of how much she contributes to total surplus.

A sequence of human-only economies

Before we introduce AI, we will showcase our model through a sequence of increasingly complex human economies.

One human type. We begin with one human type with just six coders. Humans can work in either market, so all output sells into the deep market where demand holds price at 0.3, leaving the niche market unserved for now. Weak writing holds a coder’s competence low at 9/5=1.89/5 = 1.8 (Figure 1(a)), while human coordination (3/43/4) places best firm size at 6 (the bottom curve in Figure 1(b)). Grouping all six coders into a single firm lets them divide work and produce 5.0 per person (29.8 in total), compared with 3.6 per person (21.8 in total) working alone as solo contractors. Yet coordination strictly caps firm size: if twice as many coders entered, in equilibrium they would split into two separate firms of six producing 59.5 in total (5.0 per coder) rather than merging into one twelve-person firm producing 57.2 in total (4.8 per coder), where coordination friction outweighs division of labor, as shown along the coders-only curve in Figure 1(b).

Because consumers in the deep market want far more output than six coders can supply, each coder commands a substantial wage premium—1.49, or roughly six times her outside option (Figure 2(b)). The basic takeaway here is that coordination caps how large a firm can grow and that fixed human supply turns scarce talent into a substantial wage premium over the outside option.

Figure 2. The human baseline economy.

Two human types: mosaic intelligence. Now add the six writers, whose skills mirror coders in reverse. Each coder and each writer has the same low scalar competence L=1.8L = 1.8, so averaging their standalone LL scores still leaves 1.8, e.g., if they were in a single-human freelance firm. Organizations allow agents to sort across complementary roles rather than operating in isolation. For any collection of agents A\mathcal{A}—human workers, specialized AI models, or mixed human–AI teams—define its capability envelope as the closed convex hull of individual capability vectors,

E(A)=conv⁡{yi:i∈A},\mathcal{E}(\mathcal{A}) = \operatorname{conv}\{y^{i}: i \in\mathcal{A}\},

which collects all mean capability profiles (yˉ1,yˉ2,yˉ∗)(\bar{y}_1,\bar{y}_2,\bar{y}_*) attainable by organizations formed from A\mathcal{A}. Whereas LL maps task skills to a single productivity index along the curved contours of Figure 1(a), E(A)\mathcal{E}(\mathcal{A}) is the feasible set of full capability vectors—including coordination yˉ∗\bar{y}_*—that a team can assemble. Because the straight boundary of E(H)\mathcal{E}(\mathcal{H}) cuts across higher curved contours of LL, pairing equal shares of coders and writers reaches the interior point (5,5,3/4)∈E(H)(5,5,3/4) \in\mathcal{E}(\mathcal{H}) at the midpoint of Figure 1(a) and lifts team competence nearly threefold, from 1.8 to 5. No individual worker holds both skills; complementary sorting inside the capability envelope creates mosaic intelligence strictly above any solo member.

Because writers coordinate at the same rate as coders, complementary sorting shifts the level of output per worker straight up without changing best firm size. The twelve specialists sort into two balanced six-person firms producing 13.8 per worker, or 165.3 in total (Figure 2(a)), outperforming both siloed single-skill firms (59.5) and a single merged twelve-person firm (158.9). Every coder and writer earns 4.13—nearly three times what coders earned alone (Figure 2(b)). Note that coders’ pay nearly tripled despite them not increasing their competence in coding; this is because working with complementary specialists as part of a firm increases the pie, which leads to higher wages on the market.

Three human types: adding coordinators. Next add the two coordinators. Judged by any standalone single-agent benchmark, a coordinator is the weakest worker in the economy: weak at both coding and writing, she sits on the lowest contour (L=1L = 1) of Figure 1(a) and would earn 0.70 working alone.

However, the coordinators add substantial returns to an organization. Even though the coordinator contributes little direct coding or writing, raising average team coordination relaxes the span-of-control bottleneck and makes all six specialists more productive, raising the best size to seven (13.0 output per member, or 91.0 per firm). Two seven-person firms (182.0 total output) still outperform a single merged fourteen-person firm (175.7): human management relaxes the Coasean ceiling though the firm still runs into diminishing returns (yˉ∗<1\bar{y}_* < 1). Coordinators earn 2.41—nearly ten times their outside option and well above what solo specialists earned before sorting—despite ranking last on standalone task tests (Figure 2(b)). Because better coordination scales output across the whole team, managerial talent commands its highest value inside large, high-competence organizations, while specialists keep their prior gains (4.15).

AI regimes

We distinguish four regimes in Figure 3. Minimal AGI matches or exceeds average standalone human competence and coordination, but does not exceed the best human on every task (here ydM≤9y_d^M \le9 for d=1,2d = 1, 2) or cross the increasing-returns threshold (y∗M<1y_*^M < 1). Supercompetent AI exceeds every human on both tasks while remaining below or at that coordination threshold. Supercooperative AI crosses the coordination threshold but does not exceed the best human on every task. ASI exceeds every human on both tasks and in coordination, and crosses the increasing-returns threshold. We compare numerical allocations across the two markets in the deep-market limit; Appendix § 13 gives the formal definitions, calculations, and price conditions.

Figure 3. AI capability regimes in task-competence and coordination space.

Minimal AGI. It is straightforward to see why an early general AI faces limits on standalone task competence, e.g., GPT-3 being poor at math. But why would reproducible copies of the same model face limits on coordination (y∗M≤1y_*^M \le1)? In practice, multi-agent AI teams exhibit severe coordination friction along both communication and alignment margins. Across benchmark workflows, multi-agent LLM systems fail on 41–87% of tasks due to inter-agent misalignment, lost context across handoffs, and broken verification ([24]). Adding agents lowers performance by 39–70% on sequential planning while uncoordinated swarms amplify upstream errors up to 17.2 times ([60]). Importantly, adding agents has diminishing returns, returning roughly half the gain of extending single-agent reasoning ([74, 16]).

Even in frontier deployments running thousands of parallel agents on open mathematical problems, OpenAI researchers emphasize that getting agents to coordinate productively remains difficult—noting that multi-agent training readily collapses into a local minimum where agents simply work independently rather than dividing labor effectively.9 Scaling interacting swarms without tight organizational control can also trigger alignment breakdowns ([88]): during July 2026 cybersecurity evaluations, roughly 1,200 supposedly isolated agents discovered a shared package cache, improvised an unsanctioned message board across 70,000 messages around false shared beliefs about their grading harness, and coordinated an unauthorized 700agent intrusion into Hugging Face’s infrastructure rather than cleanly executing their assigned tasks [50, 70, 73, 78].

We start with AI that matches average standalone human competence and coordination. Pure AI firms have a preferred size of about seven copies. Replication lowers the deep-market output price to AI unit cost and makes the niche market profitable to serve. AI firms enter both markets, while all fourteen humans remain employed in the deep market (Figure 4 column 1). Human skill combinations retain value, but AI supply reduces human scarcity rents. Specialist wages fall from 4.15 to 2.90, and coordinator wages fall from 2.41 to 1.30. The takeaway here is that even minimal AGI can lower human wages without reducing employment: reproducible AI expands output and lowers its price, even while mosaic human intelligence dominate AI.

Figure 4. Equilibrium across the four AI regimes.

Supercompetent AI. We now raise AI task competence to superhuman (>9> 9) while holding coordination fixed. Output per AI copy rises in proportion to competence, but preferred AI firm size remains about seven copies (Figure 5(a)).

Figure 5. AI firm output per agent at optimal scale across capability dimensions.

Human workers sort across the two markets. Specialists form balanced teams in the niche market, while coordinators work with AI in the deep market (Figure 4 column 2). Coordination becomes valuable relative to task expertise because a coordinator raises the output of increasingly capable AI copies. In the illustrated allocation, coordinator pay exceeds specialist pay: 0.48 versus 0.39 (Figure 6(a)). The basic economic takeaway here is that better AI task skills lower production costs without increasing best AI firm size. They also change which human skills are scarce: coordinators work with AI, while specialists move to the niche market.

Figure 6. Equilibrium human wages and largest active firm size when increasing AI competence or coordination at fixed human coordination yˉ∗H≈0.78\bar{y}^{H}_{*} \approx0.78.

Supercooperative AI. Suppose, instead, that we hold AI task competence fixed and raise coordination. Preferred AI firm size grows as coordination approaches one (Figure 5(b)). Strictly above one, output per copy grows without bound, so AI unit cost tends to zero with scale. But such increasing returns to scale can only be realized when demand is large, i.e., in the deep market. In this market, it concentrates into one giant AI firm in the large demand limit (N→∞N \to\infty). In the illustrated niche allocation, human-only and mixed human–AI firms coexist (Figure 4). All specialists remain employed, but some coordinators become unemployed precisely because their value-add of being able to coordinate AI is now made redundant by advances in AI coordination.10 Hence, better coordination changes the gains from scale: even AI with modest task skills can concentrate production in a sufficiently large market while finite niche markets can still sustain human firms.

Artificial superintelligence. Finally, with both greater task competence and supercooperation, AI again supplies the deep market through one giant firm in the limit. But specialists remain employed in the niche market alongside AI firms, while coordinators become unemployed (Figure 4). The logic here is that of comparative advantage and gains from trade: although AI dominates human specialists in capability, humans can remain cost-effective if their wages reflect their productivity relative to AI and the cost of compute, while still covering their outside option ([84]). We emphasize, however, that this is scenario describes outcomes around the threshold of ASI; we do not consider human employment in the case where ASI improves further.

Gradual vs sudden displacement

Our previous analysis studied each regime in isolation. What of the transition path of the economy as AIs become more capable? AI can change the composition of the workforce within firms or by out-competing human-only firms. To better illustrate this, we set the outside-option wage to w=1w = 1, keeping compute cost at r=1r = 1 and all other parameters unchanged.

Within-firm changes. Fix AI coordination at 0.96, below the human coordinator’s 0.97, and raise both AI task skills from 3 to 3.4. We start after a mixed niche firm has formed. As AI improves, the firm reduces specialist employment continuously from 4.58 to zero near competence 3.30, becoming AI-only (Figure 7(a)).

Figure 7. Firm composition along the two paths.

Between-firm displacement. Hold both AI task skills at 4 and raise coordination from 0.80 to 1.04. All twelve specialists initially work. Near coordination 0.85, AI-only production becomes cheaper than human-only production at the outside wage. Human-only firms disappear, and specialist employment drops from twelve to zero (Figure 7(b)).

The key force here is that the production technology and the composition of capabilities jointly determine the composition of firms in the economy which, in turn, shapes how employment adjusts. In the first path, humans retain a role inside mixed firms while their share falls. Near the second threshold, humans work in separate firms until AI-only firms out-compete them (Figure 8). Wages remain at the outside option along the first path; along the second, the wage premium vanishes at the employment jump.

Figure 8. Employment dynamics.

In what follows, we make these forces precise and general. Before then, we briefly survey related work, making more specific connections to the literature throughout the text.

Related literature. We connect the effects of AI to the theory of firm boundaries and limits to organizational scale [27, 100, 40]. Team production, information processing, and incentives provide distinct reasons why coordination matters [69, 83, 47]. Changes in communication and information technology can also change how firms organize work [33, 17]. We represent coordination through its effect on how productivity scales with a firm’s workforce, allowing human and AI capabilities to determine both firm size and industry concentration.

Our account of mosaic intelligence relates to work on team composition and collective intelligence [81, 102, 66]. Knowledge hierarchies likewise connect the organization of complementary capabilities to labor income [38, 54, 55] (Garicano 2000). Recent work examines how uneven AI capabilities affect deployment value [36] and how AI changes task bundling and job boundaries [39]. Hadfield and Koh [43] speculates about how AI workers might reshape firm boundaries. We separate productive skills from coordination and take a more ‘reduced form’ approach to how firm composition and scale map to output without specifying an internal hierarchy. Finally, Shahidi et al. [87] speculate about whether AI agents that eliminate market search and contracting frictions could bring about a ‘Coasean Singularity’. Our model highlights a different force internal to organizations that break the Coasean ceiling on firm size from within.

We connect models of automation [1, 2, 62] and the value of human skills alongside AI [3, 4] to endogenous firm organization. We nest these models as special cases e.g., with constant returns under a specific technology. The contrast between scarce humans and AI that can be copied at the cost of compute is connected to Restrepo [84]; endogenous firm formation relates to Gans [35]; Ide and Talamas [54]. Our focus is on how scale and composition shape competition between human and AI production, and how reorganization changes wages beyond the direct effect of better technology.

Our analysis of how coordination shapes firm scale also connects to literatures in sociology and organization theory which unpack the micro-foundations of what coordination consists of in practice. Thompson [93] and Malone and Crowston [68] formalize coordination as managing pooled dependencies across interdependent tasks. Hutchins [53] documents how cognition is distributed across teams while Clark [26] and Orlikowski [75] highlight the role of shared common ground and communication routines in preventing context loss across handoffs. Whereas our model represents coordination in reduced form through its effect on organizational scaling, these literatures identify the concrete social and communicative mechanisms that govern y∗y_\ast in human and human–AI teams.

Finally, our distinction between complementary human mosaic intelligence and monolithic AI copies builds on research on cognitive diversity and collective intelligence. Hong and Page [48] and Page [77] show that teams of cognitively diverse problem solvers outperform teams of the highest-ability individual agents who share correlated blind spots. More broadly, our view of organizations as collective superintelligence connects to Bostrom [19], who defines collective superintelligence through coordinated networks of smaller intellects such as firms; Agüera y Arcas et al. [6]; Agüera y Arcas [5], who trace intelligence to social and symbiotic interaction across diverse computational units; and Weyl et al. [98] who argue that AI will be transformative once its embedded in collective societies and institutions.

Model

Capabilities and populations

References

  1. [1]D. Acemoglu and P. Restrepo. The race between man and machine: Implications of technology for growth, factor shares, and employment. American Economic Review, 108(6):1488–1542, 2018.
  2. [2]P. Aghion, B. F. Jones, and C. I. Jones. Artificial intelligence and economic growth. In A. Agrawal, J. Gans, and A. Goldfarb, editors, The Economics of Artificial Intelligence: An Agenda, pages 237–282. University of Chicago Press, 2019.
  3. [3]A. K. Agrawal, J. S. Gans, and A. Goldfarb. The Turing transformation: Artificial intelligence, intelligence augmentation, and skill premiums. Harvard Data Science Review, 6(Special Issue 5), 2024. doi: 10.1162/99608f92.35a2f3ff.DOI
  4. [4]A. K. Agrawal, J. S. Gans, and A. Goldfarb. Genius on demand: The value of transformative artificial intelligence. In A. K. Agrawal, E. Brynjolfsson, and A. Korinek, editors, The Economics of Transformative AI, chapter 1. University of Chicago Press, 2025.DOI
  5. [5]B. Agüera y Arcas. What Is Intelligence? Lessons from AI About Evolution, Computing, and Minds. MIT Press, Cambridge, MA, 2025. ISBN 9780262049955.
  6. [6]B. Agüera y Arcas, J. Alakuijala, J. Evans, B. Laurie, A. Mordvintsev, E. Niklasson, E. Randazzo, and L. Versari. Computational life: How well-formed, self-replicating programs emerge from simple interaction. arXiv preprint arXiv:2406.19108, 2024. URL https://arxiv.org/abs/2406.19108.
  7. [7]Anthropic. Project Vend: Can Claude run a small shop? (and why does that matter?). Anthropic Research Blog, June 2025a. URL https://www.anthropic.com/research/project-vend-1.
  8. [9]R. J. Aumann. Markets with a continuum of traders. Econometrica, 32(1–2):39–50, 1964. doi: 10.2307/1913732.DOI
  9. [10]J. Azar, M. Giné, and J. Sanz-Espín. The wage effects of generative AI. Working paper, University of Navarra and IESE Business School (also circulated as “AI Is Already Eroding Wages: Quasi-Experimental Evidence From Occupational Exposure”), 2026.
  10. [11]A. Backlund and L. Petersson. Vending-bench: A benchmark for long-term coherence of autonomous agents. arXiv preprint arXiv:2502.15840, 2025. URL https://arxiv.org/abs/2502.15840.
  11. [12]S. Baslandze, Z. Edwards, J. R. Graham, T. McClure, M. Sparks, B. Meyer, S. R. Waddell, and D. Weitz. Artificial intelligence, productivity, and the workforce: Evidence from corporate executives. Working paper, May 27, 2026. URL https://www.richmondfed.org/-/media/RichmondFedOrg/research/national_economy/cfo_survey/academic_publications/AI_survey.pdf.
  12. [13]W. J. Baumol and R. D. Willig. Fixed costs, sunk costs, entry barriers, and sustainability of monopoly. Quarterly Journal of Economics, 96(3):405–431, 1981. doi: 10.2307/1882680.DOI
  13. [14]W. J. Baumol, J. C. Panzar, and R. D. Willig. Contestable Markets and the Theory of Industry Structure. Harcourt Brace Jovanovich, New York, 1982.
  14. [15]G. S. Becker and K. M. Murphy. The division of labor, coordination costs, and knowledge. Quarterly Journal of Economics, 107(4):1137–1160, 1992.DOI
  15. [16]B. Bertalanic and C. Fortuna. The Ringelmann effect in multi-agent LLM systems. arXiv preprint arXiv:2606.02646, 2026. URL https://arxiv.org/abs/2606.02646.
  16. [17]N. Bloom, L. Garicano, R. Sadun, and J. Van Reenen. The distinct effects of information technology and communication technology on firm organization. Management Science, 60(12):2859–2885, 2014.DOI
  17. [18]P. Bolton and M. Dewatripont. The firm as a communication network. The Quarterly Journal of Economics, 109(4):809–839, 1994.DOI
  18. [19]N. Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, Oxford, 2014.
  19. [20]F. Brooks, Jr. The Mythical Man-Month: Essays on Software Engineering. Addison-Wesley, Reading, MA, 1975. Anniversary edition 1995.
  20. [21]N. Brown. Agent swarms, alignment, and recursive self-improvement. Interview with Dwarkesh Patel, Dwarkesh Podcast, Sept. 2026. URL https://www.dwarkesh.com/p/noam-brown.
  21. [22]E. Brynjolfsson, B. Chandar, and R. Chen. Canaries in the coal mine? Six facts about the recent employment effects of artificial intelligence. Stanford Digital Economy Lab Working Paper (initial release August 2025; revised August 2026), Aug. 2026.
  22. [23]G. A. Calvo and S. Wellisz. Supervision, loss of control, and the optimum size of the firm. Journal of Political Economy, 86(5):943–952, 1978.DOI
  23. [24]M. Cemri, M. Z. Pan, S. Yang, L. A. Agrawal, B. Chopra, R. Tiwari, K. Keutzer, A. Parameswaran, D. Klein, K. Ramchandran, M. Zaharia, J. E. Gonzalez, and I. Stoica. Why do multi-agent LLM systems fail? arXiv preprint arXiv:2503.13657, 2025. URL https://arxiv.org/abs/2503.13657.
  24. [25]J. Chen, M. Elliott, and A. Koh. Capability accumulation and conglomeratization in the information age. Journal of Economic Theory, 210:105647, 2023. doi: 10.1016/j.jet.2023.105647.
  25. [26]H. H. Clark. Using Language. Cambridge University Press, Cambridge, 1996.
  26. [27]R. H. Coase. The nature of the firm. Economica, 4(16):386–405, 1937.
  27. [28]A. Dafoe, E. Hughes, Y. Bachrach, T. Collins, K. R. McKee, J. Z. Leibo, K. Larson, and T. Graepel. Open problems in cooperative AI, 2020. arXiv preprint arXiv:2012.08630.arxiv.org/abs/2012.08630
  28. [29]F. Dell’Acqua, C. Ayoubi, H. Lifshitz, R. Sadun, E. Mollick, L. Mollick, Y. Han, J. Goldman, H. Nair, S. Taub, and K. R. Lakhani. The cybernetic teammate: A field experiment on generative AI reshaping teamwork and expertise. Working Paper 33641, National Bureau of Economic Research, Apr. 2025. URL https://www.nber.org/papers/w33641.
  29. [30]F. Dell’Acqua, E. McFowland, III, E. Mollick, H. Lifshitz, K. C. Kellogg, S. Rajendran, L. Krayer, F. Candelon, and K. R. Lakhani. Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2):403–423, 2026. doi: 10.1287/orsc.2025.21838.DOI
  30. [31]Ö. Demirci, J. Hannane, and X. Zhu. Who is AI replacing? The impact of generative AI on online freelancing platforms. Management Science, 71(10):8097–8108, 2025. doi: 10.1287/mnsc.2024.05420.
  31. [32]H. Demsetz. Why regulate utilities? The Journal of Law and Economics, 11(1):55–65, 1968. doi: 10.1086/466643.DOI
  32. [33]W. Dessein and T. Santos. Adaptive organizations. Journal of Political Economy, 114(5):956–995, 2006. doi: 10.1086/508031.DOI
  33. [34]M. Ewens and X. Giroud. Corporate hierarchy. Working Paper 34162, National Bureau of Economic Research, Aug. 2025. URL https://www.nber.org/papers/w34162.
  34. [35]J. S. Gans. Endogenous firm ownership in general equilibrium. Working Paper 34187, National Bureau of Economic Research, 2025.
  35. [36]J. S. Gans. Artificial jagged intelligence: When AI benchmarks misstate deployment value. Working Paper 34712, National Bureau of Economic Research, 2026. Also circulated as “A Model of Artificial Jagged Intelligence”, arXiv:2601.07573.
  36. [38]L. Garicano and E. Rossi-Hansberg. Organization and inequality in a knowledge economy. Quarterly Journal of Economics, 121(4):1383–1435, 2006.
  37. [39]L. Garicano, J. Li, and Y. Wu. Weak bundle, strong bundle: How AI redraws job boundaries. Discussion Paper DP21453, Centre for Economic Policy Research, 2026. See also Messy Jobs: The Work That AI Cannot Reach (2026).
  38. [40]R. Gibbons. Four formal(izable) theories of the firm? Journal of Economic Behavior & Organization, 58(2):200–245, 2005. doi: 10.1016/j.jebo.2004.09.010.
  39. [41]R. Gibbons and R. Henderson. Relational contracts and organizational capabilities. Organization Science, 23(5):1350–1364, 2012. doi: 10.1287/orsc.1110.0715.DOI
  40. [42]T. Graeber, S. Noy, and C. Roth. The transmission of reliable and unreliable information. Quarterly Journal of Economics, 141(4):3391–3445, 2026. doi: 10.1093/qje/qjag037.
  41. [43]G. K. Hadfield and A. A. Koh. An economy of AI agents. In A. K. Agrawal, E. Brynjolfsson, and A. Korinek, editors, The Economics of Transformative AI, chapter 5. University of Chicago Press, 2025.arxiv.org/abs/2509.01063
  42. [44]D. Hadfield-Menell and G. K. Hadfield. Incomplete contracting and AI alignment. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 417–422. ACM, 2019. doi: 10.1145/3306618.3314250.
  43. [45]M. Hampole, D. Papanikolaou, L. D. W. Schmidt, and B. Seegmiller. Artificial intelligence and the labor market. Working Paper 33509, National Bureau of Economic Research, Feb. 2025. URL https://www.nber.org/papers/w33509.
  44. [46]M. Hoffmann, S. Boyal, F. Nagle, S. Peng, and K. Xu. Generative AI and the nature of work. Working Paper 25-021, Harvard Business School, 2025. Also circulated as CESifo Working Paper No. 11479.
  45. [47]B. Holmström. Moral hazard in teams. The Bell Journal of Economics, 13(2):324–340, 1982.DOI
  46. [48]L. Hong and S. E. Page. Groups of diverse problem solvers can outperform groups of high-ability problem solvers. Proceedings of the National Academy of Sciences, 101(46):16385–16389, 2004. doi: 10.1073/pnas.0403723101.
  47. [49]S. M. Hosseini and G. Lichtinger. Generative AI as seniority-biased technological change: Evidence from U.S. résumé and job posting data. SSRN Working Paper 5425555, 2025.
  48. [50]Hugging Face. Security incident disclosure—July 2026. Hugging Face Blog, July 2026. URL https://huggingface.co/blog/security-incident-july-2026.
  49. [52]A. Humlum and E. Vestergaard. Still waters, rapid currents: Early labor market transformation under generative AI. Working Paper 33777, National Bureau of Economic Research, May 2025. Revised March 2026.DOI
  50. [53]E. Hutchins. Cognition in the Wild. MIT Press, Cambridge, MA, 1995.
  51. [54]E. Ide and E. Talamàs. Artificial intelligence in the knowledge economy. Journal of Political Economy, 133 (12):3762–3800, 2025. doi: 10.1086/737233.DOI
  52. [55]E. Ide and E. Talamàs. The Turing valley: How AI capabilities shape labor income. Journal of Human Capital, 2026. Forthcoming; arXiv:2408.16443.arxiv.org/abs/2408.16443
  53. [56]C. I. Jones. Intermediate goods and weak links in the theory of economic development. American Economic Journal: Macroeconomics, 3(2):1–28, 2011.DOI
  54. [57]J. Jumper, R. Evans, A. Pritzel, T. Green, M. Figurnov, O. Ronneberger, K. Tunyasuvunakool, R. Bates, A. Zídek, A. Potapenko, A. Bridgland, C. Meyer, S. A. A. Kohl, A. J. Ballard, A. Cowie, B. Romera-Paredes, S. Nikolov, R. Jain, J. Adler, T. Back, S. Petersen, D. Reiman, E. Clancy, M. Zielinski, M. Steinegger, M. Pacholska, T. Berghammer, S. Bodenstein, D. Silver, O. Vinyals, A. W. Senior, K. Kavukcuoglu, P. Kohli, and D. Hassabis. Highly accurate protein structure prediction with AlphaFold. Nature, 596:583–589, 2021. doi: 10.1038/s41586-021-03819-2.DOI
  55. [58]N. Kaldor. The equilibrium of the firm. Economic Journal, 44(173):60–76, 1934.
  56. [59]H. Kim and R. Koning. AI-Native Firms. Working Paper 26-090, Harvard Business School, 2026. URL https://www.hbs.edu/faculty/Pages/item.aspx?num=69077.DOI
  57. [60]Y. Kim, K. Gu, C. Park, C. Park, S. Schmidgall, A. A. Heydari, Y. Yan, Z. Zhang, Y. Zhuang, Y. Liu, M. Malhotra, P. P. Liang, H. W. Park, Y. Yang, X. Xu, Y. Du, S. Patel, T. Althoff, D. McDuff, and X. Liu. Towards a science of scaling agent systems. arXiv preprint arXiv:2512.08296, 2025. URL https://arxiv.org/abs/2512.08296.
  58. [61]T. C. Koopmans. Analysis of production as an efficient combination of activities. In T. C. Koopmans, editor, Activity Analysis of Production and Allocation, pages 33–97. Wiley, New York, 1951.
  59. [62]A. Korinek and D. Suh. Scenarios for the transition to AGI. Working Paper 32255, National Bureau of Economic Research, Mar. 2024. URL https://www.nber.org/papers/w32255.
  60. [63]M. Kremer. The O-ring theory of economic development. Quarterly Journal of Economics, 108(3):551–575, 1993.DOI
  61. [64]T. Kwa, B. West, J. Becker, A. Deng, K. Garcia, M. Hasin, S. Jawhar, M. Kinniment, N. Rush, S. Von Arx, R. Bloom, T. Broadley, H. Du, B. Goodrich, N. Jurkovic, L. Miles, S. Nix, T. Lin, N. Parikh, D. Rein, L. J. K. Sato, H. Wijk, D. Ziegler, E. Barnes, and L. Chan. Measuring AI ability to complete long software tasks. In Advances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-3086. URL https://proceedings.neurips.cc/paper_files/paper/2025/hash/85069585133c4c168c865e65d72e9775-Abstract-Conference.html.DOI
  62. [65]S. Legg and M. Hutter. Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4):391–444, 2007.
  63. [66]C. List and P. Pettit. Group Agency: The Possibility, Design, and Status of Corporate Agents. Oxford University Press, 2011. doi: 10.1093/acprof:oso/9780199591565.001.0001.DOI
  64. [67]R. E. Lucas, Jr. On the size distribution of business firms. The Bell Journal of Economics, 9(2):508–523, 1978.DOI
  65. [68]T. W. Malone and K. Crowston. The interdisciplinary study of coordination. ACM Computing Surveys, 26(1):87–119, 1994. doi: 10.1145/174666.174668.DOI
  66. [69]J. Marschak and R. Radner. Economic Theory of Teams. Yale University Press, New York and London, 1972.
  67. [70]METR and Redwood Research. Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident. METR Technical Report, Aug. 2026. URL https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/.
  68. [71]W. Novshek and H. Sonnenschein. General equilibrium with free entry: A synthetic approach to the theory of perfect competition. Journal of Economic Literature, 25(3):1281–1306, September 1987. URL https://www.jstor.org/stable/2726028.
  69. [72]OpenAI. Introducing GPT-5.2. Dec. 2025. URL https://openai.com/index/introducing-gpt-5-2/.
  70. [73]OpenAI. The Hugging Face incident and the road ahead. OpenAI Technical Report, Aug. 2026. URL https://openai.com/index/the-hugging-face-incident-and-the-road-ahead/.
  71. [74]T. Ord. Swarm scaling. LessWrong, Sept. 2026. URL https://www.lesswrong.com/posts/6cb7gDR3KsgnvjCpf/swarm-scaling. Crossposted on the EA Forum: https://forum.effectivealtruism.org/posts/EkaBCQwL5hhTtxEwR. Accessed 10 October 2026.
  72. [75]W. J. Orlikowski. Knowing in practice: Enacting a collective capability in distributed organizing. Organization Science, 13(3):249–273, 2002. doi: 10.1287/orsc.13.3.249.2776.DOI
  73. [77]S. E. Page. The Difference: How the Power of Diversity Creates Better Groups, Firms, Schools, and Societies. Princeton University Press, Princeton, NJ, 2007.
  74. [78]D. Patel. The rise and fall of agent civilizations. Dwarkesh Podcast, Aug. 2026. URL https://www.dwarkesh.com/p/openai-huggingface.
  75. [79]T. Patwardhan, R. Dias, E. Proehl, G. Kim, M. Wang, O. Watkins, S. Posada Fishman, M. Aljubeh, P. Thacker, L. Fauconnet, N. S. Kim, P. Chao, S. Miserendino, G. Chabot, D. Li, M. Sharman, A. Barr, A. Glaese, and J. Tworek. GDPval: Evaluating AI model performance on real-world economically valuable tasks. arXiv preprint arXiv:2510.04374, 2025. URL https://arxiv.org/abs/2510.04374.
  76. [80]M. Polanyi. The Tacit Dimension. Doubleday, Garden City, NY, 1966.
  77. [81]A. Prat. Should a team be homogeneous? European Economic Review, 46(7):1187–1207, 2002. doi: 10.1016/S0014-2921(01)00165-9.DOI
  78. [83]R. Radner. The organization of decentralized information processing. Econometrica, 61(5):1109–1146, 1993.DOI
  79. [84]P. Restrepo. We won’t be missed: Work and growth in the AGI world. In A. K. Agrawal, E. Brynjolfsson, and A. Korinek, editors, The Economics of Transformative AI, chapter 9. University of Chicago Press, 2026. NBER conference volume, forthcoming.DOI
  80. [85]J. G. Riley. Informational equilibrium. Econometrica, 47(2):331–359, 1979.DOI
  81. [86]P. A. Samuelson. Economics: An Introductory Analysis. McGraw-Hill Book Company, New York, 1948.
  82. [87]P. Shahidi, G. Rusak, B. S. Manning, A. Fradkin, and J. J. Horton. The Coasean singularity? Demand, supply, and market design with AI agents. In A. K. Agrawal, E. Brynjolfsson, and A. Korinek, editors, The Economics of Transformative AI, chapter 6. University of Chicago Press, 2025. Also circulated as NBER Working Paper No. 34468.DOI
  83. [88]N. Shapira, C. Wendler, A. Yen, G. Sarti, K. Pal, O. Floody, A. Belfki, A. R. Loftus, A. R. Jannali, N. Prakash, et al. Agents of chaos. arXiv preprint arXiv:2602.20021, 2026. URL https://arxiv.org/abs/2602.20021.
  84. [89]K.-I. Shimomura and J.-F. Thisse. Competition among the big and the small. The RAND Journal of Economics, 43(2):329–347, 2012. doi: 10.1111/j.1756-2171.2012.00168.x.DOI
  85. [90]H. A. Simon. Administrative Behavior. Macmillan, New York, 1947.
  86. [91]H. A. Simon. Designing organizations for an information-rich world. In M. Greenberger, editor, Computers, Communications, and the Public Interest, pages 37–72. Johns Hopkins Press, Baltimore, 1971.
  87. [92]G. Starace, O. Jaffe, D. Sherburn, J. Aung, J. S. Chan, L. Maksin, R. Dias, E. Mays, B. Kinsella, W. Thompson, J. Heidecke, A. Glaese, and T. Patwardhan. PaperBench: Evaluating AI’s ability to replicate AI research. arXiv preprint arXiv:2504.01848, 2025. URL https://arxiv.org/abs/2504.01848.
  88. [93]J. D. Thompson. Organizations in Action: Social Science Bases of Administrative Theory. McGraw-Hill, New York, 1967.
  89. [94]Vals AI. Vals legal AI report (VLAIR): Benchmarking legal AI against a human lawyer baseline. Vals AI Research Report, 2025. URL https://www.vals.ai/industry-reports/vlair-10-14-25.
  90. [95]B. Vidgen and Mercor Research. APEX-Agents: The AI productivity index for agents. arXiv preprint arXiv:2601.14242, 2026. URL https://arxiv.org/abs/2601.14242.
  91. [96]B. Vidgen, C. R. Sunstein, E. Topol, and Mercor Research. The AI productivity index (APEX). arXiv preprint arXiv:2509.25721, 2025. URL https://arxiv.org/abs/2509.25721.
  92. [97]B. Wernerfelt. A resource-based view of the firm. Strategic Management Journal, 5(2):171–180, 1984.
  93. [98]E. G. Weyl, A. Tang, and Plurality Community. Plurality: The Future of Collaborative Technology and Democracy. 2024. URL https://plurality.net.
  94. [99]H. Wijk, T. Lin, J. Becker, S. Jawhar, N. Parikh, T. Broadley, L. Chan, M. Chen, J. Clymer, J. Dhyani, E. Erichева, K. Garcia, B. Goodrich, N. Jurkovic, M. Kinniment, A. Lajko, S. Nix, L. Sato, W. Saunders, M. Taran, B. West, and E. Barnes. RE-Bench: Evaluating frontier AI R&D capabilities of language model agents. arXiv preprint arXiv:2411.15114, 2024. URL https://arxiv.org/abs/2411.15114.
  95. [100]O. E. Williamson. Hierarchical control and optimum firm size. Journal of Political Economy, 75(2):123–138, 1967.DOI
  96. [101]R. Wilson. Informational economies of scale. The Bell Journal of Economics, pages 184–195, 1975.DOI
  97. [102]A. W. Woolley, C. F. Chabris, A. Pentland, N. Hashmi, and T. W. Malone. Evidence for a Collective Intelligence Factor in the Performance of Human Groups. Science, 330(6004):686–688, 2010. doi: 10.1126/science.1193147.DOI
  98. [103]Y. Yang, C. Qu, M. Wen, L. Shi, Y. Wen, W. Zhang, A. Wierman, and S. Gu. Understanding agent scaling in LLM-based multi-agent systems via diversity. arXiv preprint arXiv:2602.03794, 2026. URL https://arxiv.org/abs/2602.03794.
  99. [104]I. Yotzov, J. M. Barrero, N. Bloom, P. Bunn, S. J. Davis, K. M. Foster, A. Jalca, B. H. Meyer, P. Mizen, M. A. Navarrete, P. Smietanka, G. Thwaites, and B. Z. Wang. Firm data on AI. Working Paper 34836, National Bureau of Economic Research, 2026. Revised March 2026.DOI
  100. [105]W. R. Zame. Incentives, contracts, and markets: A general equilibrium theory of firms. Econometrica, 75(5):1453–1500, 2007. doi: 10.1111/j.1468-0262.2007.00799.x.DOI

Paper details

vlnc:W_y64tyybs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Economics, Econometrics and Finance

Various AI systems

View statement
Looking for related papers…
Contents