Cognitive labor substitution, organizational hyper-flattening and the emergence of new operational risks
Economic theory long held as given that technological progress increased the marginal productivity of workers without ever threatening their monopoly over complex cognitive tasks. This complementarity hypothesis, formalized by Autor, Levy and Murnane as early as 2003 in what the literature came to know as the ALM model, posited an automation boundary deemed uncrossable: machines could encode explicit rules and process structured data, but orchestration, judgment and coordination remained the irreducible preserve of the skilled human worker.1 Autor himself termed this barrier the "Polanyi paradox": we always know more than we can say, and this tacit knowledge protected the intellectual worker from algorithmic substitution.
The emergence of Agentic Artificial Intelligence in the mid-2020s shatters this consensus. This is no longer a passive generative AI that produces a response to a prompt and then stops, waiting for the human to orchestrate the next step. Agentic AI, whose most documented iterative cognitive architecture is the ReAct framework formalized by Yao et al. in 2022, autonomously chains a reasoning loop: it decomposes a complex objective into subtasks, selects and activates external tools such as web browsing or business APIs, analyzes the results of its own actions and self-corrects without human intervention, until the initial directive is accomplished.2 This shift, from oracle mode to agentic mode, is precisely what the ALM model could not anticipate: the machine no longer assists the intellectual worker, it internalizes the complete workflow that was until then their monopoly.
The rupture is confirmed by the most recent empirical data. Eloundou et al. showed in 2023, in a landmark study, that it is now the highest-earning professions with the greatest cognitive component that exhibit the greatest exposure to algorithmic substitutability, spectacularly reversing the historical prediction of the employment polarization model.3 The Polanyi paradox was not resolved theoretically; it was circumvented empirically by probabilistic deep learning, which does not need explicit instructions to generate expert-quality reasoning.
Before measuring the organizational impact of agentic AI, it is essential to reclassify its economic nature. Classical literature unambiguously placed computing in the category of technical capital: the acquisition of servers, software licenses or infrastructure constituted capital expenditure, a CAPEX generating fixed costs amortized over several years. The transition to SaaS initiated a first shift toward operating expenses (OPEX), but the cost remained fixed in the short term for the company, regardless of usage intensity.
The commercialization of large language models via programming interfaces (APIs), in particular by OpenAI and Anthropic from 2023 onward, marks a radical break: pricing now occurs at the microscopic usage level, in tokens, that is, in processed semantic units. Technological expenditure thus shifts from a fixed cost to a purely variable charge, strictly proportional to the quantity of intellectual tasks executed. In the event of a negative demand shock, the algorithmic cost drops instantly to zero, something no human employment contract can replicate.
This mutation is not merely accounting in nature. Lazear theorized it as early as 2000 for human labor: piece-rate compensation aims to perfectly align the cost of labor with the marginal productivity of the executor, eliminating the idle time inherent to hourly wages.4 This is exactly the structure that per-token billing reproduces: the firm no longer pays for the intellectual availability of an executive, it remunerates the direct and punctual execution of the cognitive task, via an instantaneous commercial contract that completely bypasses the legal rigidities of labor law. The token is the piece-rate wage of agentic AI.
The absolute elasticity of algorithmic supply only reveals its full significance when confronted with the pricing reality of labor markets. Available data converges toward a univocal diagnosis: the cost differential between human execution and agentic execution on skilled cognitive tasks is of an order of magnitude that makes the comparison almost absurd.
To formalize this value destruction and test its robustness against regulatory frictions, a dynamic microeconomic simulator was developed. It projects the processing cost of a batch of fifty audit reports (approximately 500,000 input tokens) across three architectures: the human floor paid at the gross minimum wage (160 €), the recourse to the Cloud API indexed on the most expensive market models such as GPT-5.5 or Claude Opus at their spring 2026 pricing (3.02 €), and local sovereign deployment on an amortized RTX 4090 workstation (0.17 €). The simulator integrates comparative error rates, all market parameters and calculates the compensatory tax the state would need to impose for human labor to regain competitiveness, yielding a punitive tax rate exceeding 5,000%.
The central result of the modeling does not lie in the raw pricing crush, already spectacular in itself. Its true analytical strength lies in the integration of non-quality risk. Even when parameterizing the simulator with a deliberately inflated algorithmic hallucination rate of 8% versus 5% errors for a tired human, and integrating a repair premium per error, the machine's profitability remains overwhelming. This asymmetry between speed, cost and residual reliability constitutes the core of the firm's new vulnerability, which we analyze in the third section.
The collapse of marginal cognitive cost is not limited to an optimization of operating expenses; it calls into question the very foundation of the firm's existence. In his seminal 1937 article, Ronald Coase posed an apparently trivial question: if the free market is the most efficient allocation mechanism, why do we observe the formation of hierarchical organizations? The answer lay in transaction costs. Systematically resorting to the market for each micro-task generates costly frictions: search costs, negotiation costs and contract drafting costs. The firm thus arises from the necessity of internalizing these transactions under a hierarchical authority.5 Williamson subsequently formalized that the optimal size of a firm stabilizes at the point where the cost of organizing one additional transaction internally equals the cost of that same transaction on the open market.
Agentic AI operates on this theoretical frontier with unprecedented brutality. Contemporary multi-agent architectures, for which the ChatDev project at Tsinghua University offers the most accomplished quantitative demonstration, show that autonomous agents can design, code, test and document software end-to-end in less than seven minutes at an execution cost below one dollar.6 The agents exchange without psychological friction, without ego, without semantic loss, without synchronization meetings. Formally, if we denote by θ the parameter of internal coordination friction, agentic AI acts as a technological shock compressing θ toward zero.
As soon as the marginal cost of internal coordination collapses, the economic justification for maintaining a mass of human capital to ensure operational fluidity disappears. The optimal size of the firm contracts, not by managerial decision, but by pure cost-minimization logic.
This obsolescence of internal human coordination mechanically calls into question the hierarchical structure that supported it. Alfred Chandler demonstrated in 1977 that the "visible hand" of management had replaced the "invisible hand" of the market in orchestrating the complex productive processes of the large industrial enterprise.7 The manager became the central processor of information, translating strategic orientations into descending operational orders and synthesizing activity data into ascending reports. This intermediation was valuable because it compensated for a fundamental biological constraint: the bounded rationality of the single leader, whose Span of Control is intrinsically limited by their cognitive capacities.
Agentic AI removes this constraint. A system capable of monitoring, coordinating and allocating resources across thousands of simultaneous tasks without decision fatigue renders intermediate strata superfluous. McKinsey Global Institute estimates that generative AI can automate more than 70% of the time managers currently spend collecting and processing data. Eloundou et al. document that approximately 80% of senior executives' tasks are exposed to language models.8 The phenomenon of delayering or hyper-flattening, which Rajan and Wulf began documenting as early as 2006 in the wake of the computing revolution, here reaches its logical paroxysm.
This double collapse, Coasian on one side and Chandlerian on the other, makes possible the emergence of what this research conceptualizes as the "Unipersonal Empire" or "Zero-Employee Firm." This is not an absolute empirical reality in which human labor has entirely disappeared, but rather an ideal-type in the Weberian sense, a theoretical asymptote allowing us to observe how classical organizational theories behave when marginal coordination cost is pushed toward zero. In this structure, a sole director-architect completely decouples the exponential growth of their output from the volume of their payroll. The already visible examples of this trajectory are telling: Midjourney generates hundreds of millions of dollars in revenue with just a few dozen collaborators; the acquisition of WhatsApp for 19 billion dollars with 55 employees remains the archetype of this decoupling between valuation and headcount.
The promise of the Zero-Employee Firm rests on a substitution presented as perfect: replacing a fallible and costly human workforce with a rational and scalable architecture of autonomous agents. But this eviction of the human factor generates a deep organizational paradox that agency theory allows us to clarify precisely. The classical agency relationship between a principal and their employee suffers from an information asymmetry coupled with a divergence of interests: the employee has their own objectives (career, effort aversion, leisure maximization) which diverge from those of the leader and produce moral hazard.9
The algorithmic agent neutralizes this interest divergence: it does not unionize, does not conceal its effort, has no agenda of its own. But it introduces an information asymmetry of a radically new nature: the inevitable opacity of deep neural networks. The human principal is now incapable of retracing the logical pathway taken by the agent in high-dimensional vector spaces to reach its decision. The so-called "black box" phenomenon is not an imperfection correctable by better engineering; it is a structural property of probabilistic architectures. Algorithmic hallucination (the production of false information generated with a high degree of mathematical certainty) is the precise contemporary equivalent of the residual loss theorized in the agency cost equation.
Beyond the individual opacity of each agent, it is the collective structure of the agentic architecture that harbors the most serious risk. In a classical firm, the risk of operational error is fundamentally idiosyncratic. The failure of a collaborator remains a localized event; by analogy with modern portfolio theory, a human team benefits from a diversification effect across its cognitive profiles. The probability that an entire department simultaneously commits the same misjudgment is statistically negligible: the law of large numbers protects the organization's overall balance sheet.
The shift toward the Zero-Employee Firm annihilates this statistical protection. By replacing the diversity of human capital with a centralized algorithmic architecture resting on the same foundation model, the leader substitutes dispersed risk for a risk of absolute concentration. If this model exhibits a latent analytical bias or a reasoning flaw, the failure is no longer isolated: it is replicated instantaneously and systematically by every agent in the company. Stanford's Center for Research on Foundation Models has theorized this "homogenization" as the creation of an internal Single Point of Failure.10 Perrow had formalized the same logic for highly coupled industrial systems: in such a system, serious accidents become not only possible but "normal" and inevitable.
In the Unipersonal Empire, the output produced by a first agent immediately becomes the input of the second execution agent, with no human buffer space to interrupt the flow. The error undergoes perfect correlation and amplifies in cascade: what was merely a local probabilistic drift metamorphoses in milliseconds into a legally binding contractual commitment. The Flash Crash of 2010, in which high-frequency trading algorithms destroyed billions in market capitalization before any human intervention, constitutes the most precise analogy of what this dynamic produces at scale in an agentic business context.
A third vulnerability adds to the first two, external in nature. The director-architect of the Zero-Employee Firm readily perceives themselves as a sovereign entity, freed from employment and its contractual rigidities. But this apparent independence masks a translation of the subordination relationship: the company has not eliminated dependence on its factors of production; it has traded its dependence on human capital for an absolute technological drip-feed from APIs and Cloud infrastructure.
For an agentic architecture to be performant, the leader must invest heavily in highly idiosyncratic intangible capital: ultra-calibrated prompts, complex workflows, models finely tuned to their business niche. Yet if the provider unilaterally modifies the architecture of their neural network, the entirety of this investment instantly loses its use value. Williamson theorized this dynamic as the "fundamental transformation": an initially competitive commercial relationship mutates into a bilateral monopoly situation as soon as switching costs become prohibitive.11 Klein, Crawford and Alchian formalized the consequence: the supplier acquires a hold-up power enabling them to extract economic rent from the captive firm, via arbitrary modification of per-token pricing, the brutal deprecation of a model, or the censoring of certain sensitive business queries in the name of opaque internal policies.
The convergence of these internal and external vulnerabilities leads to a question of institutional survival: faced with a firm exposed to instantaneous systemic accidents, how can law and the insurance market apprehend and price its failures? The current answer is troubling, for it highlights the structural incompatibility of the Zero-Employee Firm with the foundations of actuarial science.
Frank Knight established in 1921 a fundamental epistemological distinction between "risk," a situation in which the probabilities of future outcomes are known or calculable, and "uncertainty," a situation in which the very distribution of probabilities is totally unknown and incalculable. Actuarial science can only price risk.12 Yet the Unipersonal Empire structurally plunges into the domain of pure uncertainty, for three cumulative reasons.
First, it has no loss history on which to base pricing; agentic technology is too recent. Second, and more seriously, its potential failures depend on algorithms whose neural weights are permanently updated by API providers, so that the execution environment changes faster than actuaries can collect data from it. Third, Wei et al. documented that large language models develop "emergent capabilities," unpredictable stochastic behaviors they were not explicitly programmed to execute, manifesting suddenly beyond a certain scale threshold.13 As soon as an agent begins to reason according to patterns not anticipated by its creators, it becomes impossible to assign a mathematical probability to the occurrence of an interpretation error. Algorithmic losses do not follow a normal distribution; they follow the laws of Taleb's "Extremistan," where highly improbable events concentrate infinite destruction potential.
To bridge this legal chasm, the rational reflex of the director-architect is to turn to the risk transfer market, by taking out a Professional Liability policy. This attempt runs into a structural double refusal.
On the AI providers' side, contractual incompleteness is resolved by radical asymmetry. Grossman and Hart established that a contract is intrinsically incomplete when it is impossible for the parties to anticipate all future contingencies. The contract linking the Zero-Employee Firm to an API provider is the absolute archetype of this incompleteness. The technological oligopoly's response is to impose via its terms of service a total risk transfer: models are provided "as is," with absolute disclaimers of liability in the event of an erroneous decision. The provider captures virtually all of the technological rent while purging from its balance sheet any legal liability associated with the cognitive output.
On the insurers' side, the response is symmetrical. Biener, Eling and Wirfs had already documented in 2015 the market's systemic difficulty in modeling and absorbing the classical cyber risk of human enterprises, due to a lack of data and an excessively high accumulation risk.14 Insuring a firm entirely driven by an open-loop AI network, capable of committing the company's balance sheet across thousands of transactions per millisecond without human verification, far exceeds the capital absorption capacity of global reinsurers. The silent cyber exclusion clauses, which LMA 21-042 precisely sought to clean up, aim to exclude technological risks not explicitly priced; and autonomous unsupervised algorithmic execution is precisely what these clauses refuse to cover.
The leader of the Unipersonal Empire thus finds themselves legally isolated, bearing on their own equity capital alone the risk generated by their machine. Lemley and Casey recalled that AI is by nature judgment-proof, unseizable and unaccountable: having no legal personality, the algorithm cannot be sued. The leader becomes the sole Locus of Intent, the only bearer of the moral and legal risk of their fleet of agents' decisions. The institutional survival of the Unipersonal Empire will therefore depend on the emergence of alternative mechanisms: parametric insurance indexed on cryptographic oracles, dedicated reinsurance captives for systemic agentic risk, or self-insurance vehicles requiring the immobilization of massive provisions on equity.
The analysis conducted in this article outlines the contours of a fundamental anomaly in the history of organizational theory. Agentic AI brilliantly resolves the problem of human frictions theorized by Coase and made visible by Chandler; it has not abolished risk for all that, it has mutated its nature. By evacuating the social and cognitive shock absorbers of the human hierarchy, the firm has rid itself of its traditional agency costs, but has also destroyed its natural resiliences. The accident ceases to be an idiosyncratic failure and becomes a systemic crisis, instantaneous and uninsurable under current legal conditions.
The institutional constraint of the EU AI Act, which mandates the forced maintenance of human supervision in the decision loop (Human-in-the-loop) for critical processes, acts as a societal shock absorber. It slows hyper-flattening by obliging the company to retain residual human strata, no longer for their productive efficiency, but for their strict function as legal and moral guarantors. The tension between the economic logic of hyper-flattening and the regulatory logic of human control constitutes precisely the institutional fault line of the coming years.
The waves of mass layoffs currently striking the middle management layers of organizations such as Meta, Microsoft or Oracle do not constitute mere cyclical adjustments: they are the first empirical signals of the hyper-flattening theorized here. The question that future research will need to resolve lies in the market's capacity to reintegrate the Knightian uncertainty inherent in algorithms into the sphere of probabilizable risk. To invent the infallible firm is to condemn oneself to reinventing liability and insurance.
1. Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change. Quarterly Journal of Economics, 118(4); Autor, D. H. (2015). Why are there still so many jobs? Journal of Economic Perspectives, 29(3).
2. Yao, S., Zhao, J., Yu, D., et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv preprint arXiv:2210.03629.
3. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023). GPTs are GPTs: An early look at the labor market impact potential of large language models. arXiv preprint arXiv:2303.10130.
4. Lazear, E. P. (2000). Performance pay and productivity. American Economic Review, 90(5), 1346-1361.
5. Coase, R. H. (1937). The nature of the firm. Economica, 4(16), 386-405; Williamson, O. E. (1981). The economics of organization. American Journal of Sociology, 87(3).
6. Qian, C., et al. (2023). Communicative Agents for Software Development. arXiv preprint arXiv:2307.07924. Tsinghua University.
7. Chandler, A. D. (1977). The Visible Hand: The Managerial Revolution in American Business. Harvard University Press.
8. McKinsey Global Institute (2023). The economic potential of generative AI; Eloundou et al. (2023), op. cit.
9. Jensen, M. C., & Meckling, W. H. (1976). Theory of the firm: Managerial behavior, agency costs and ownership structure. Journal of Financial Economics, 3(4), 305-360.
10. Bommasani, R., Hudson, D. A., et al. (2021). On the Opportunities and Risks of Foundation Models. Stanford CRFM, arXiv preprint.
11. Williamson, O. E. (1985). The Economic Institutions of Capitalism. Free Press; Klein, B., Crawford, R. G., & Alchian, A. A. (1978). Vertical Integration, Appropriable Rents. Journal of Law and Economics, 21(2).
12. Knight, F. H. (1921). Risk, Uncertainty and Profit. Houghton Mifflin.
13. Wei, J., Tay, Y., Bommasani, R., et al. (2022). Emergent Abilities of Large Language Models. Transactions on Machine Learning Research.
14. Biener, C., Eling, M., & Wirfs, J. H. (2015). Insurability of Cyber Risk: An Empirical Analysis. Geneva Papers on Risk and Insurance, 40(1), 131-158.
The AI Act is a safety code, not a liability regime. Yet the withdrawal of the dedicated directive has made it, by default, the keystone of liability for artificial intelligence, and compliance with the regulation has become the new frontier of insurability.
Parametric insurance promises to replace the slowness of loss adjustment with the speed of measurement. This elegance has a price often left unspoken: by substituting an index for actual damage, it trades the friction of adjustment for two new dependencies, basis risk and trust in the oracle.
Solvency II rests on two postulates inherited from classical actuarial science, the ability to diversify weakly correlated risks and the ability to estimate their distribution from historical data. Cyber risk contradicts both.
The regulator asks a simple question, how much money must you hold to survive your worst year in two centuries. For a book of hurricanes, the actuary answers without flinching. For a book of ransomware, the room falls silent.
A structure falls in a town you have never heard of, and a retiree's savings in Tokyo wobble that same month. A hidden wire runs between the two, and that wire is reinsurance.