In 1937, Ronald Coase asked a question that seemed almost too obvious to be worth asking. Why do firms exist? Markets can coordinate economic activity through the price mechanism, so why do people not simply transact with one another directly rather than organising themselves into the hierarchical, rule-bound structures we call companies? Coase's answer was that transactions have costs, costs of searching for information, of negotiating agreements, of monitoring performance and enforcing contracts, and that firms exist because, under certain conditions, they can coordinate activity more cheaply than markets can (Coase, 1937).
That insight became one of the most productive ideas in twentieth-century economics. It also carried a predictive implication that has proven reliable across successive waves of technological change. As transaction costs fall, as communications improve, as information becomes cheaper to acquire and process, the optimal size of firms and the structure of industries will change. The telegraph, the telephone, the computer, and the internet each reshaped the boundaries of the firm in ways that Coase's framework helped to explain.
AI agents represent the next, and arguably the most consequential, reduction in transaction costs in economic history. These are autonomous systems capable of executing complex, multi-step tasks across multiple domains, doing things that previously required human coordination: research, analysis, drafting, negotiation, project management, customer interaction.
As these capabilities develop, they challenge not merely the efficiency of existing organisational forms but the very categories we use to think about what work is, who does it, and how economic value is created and distributed. This article argues that the Coasean implications of agent proliferation are real and significant, but that they are subject to a constraint that much current commentary neglects, namely the extraordinary computational and energy resources that frontier AI systems require. It also argues that the access question, which is more nuanced than it first appears, will determine how evenly the benefits of that transformation are distributed. The article examines each of these dimensions in turn before drawing out the institutional implications.
The Coasean Logic of Agent Proliferation
Shahidi, Rusak, Manning, Fradkin and Horton (2025), in a working paper from Harvard and MIT, offer one of the most rigorous treatments of this transition yet published. Their framing, which they term the Coasean Singularity, argues that AI agents are poised to transform digital markets by dramatically reducing transaction costs. The activities that comprise those costs, learning prices, negotiating terms, writing contracts, monitoring compliance, are precisely the tasks that AI agents can potentially perform at very low marginal cost. Once agents can execute these functions reliably, the traditional make-or-buy boundaries that define firm organisation are subject to fundamental renegotiation (Shahidi et al., 2025).
The implied structural consequence is a progressive fragmentation of the integrated firm, with smaller, specialised human organisations augmented by AI replacing the large corporate structures of the twentieth century. This fragmentation thesis exists in tension with an equally plausible concentration thesis. Acemoglu and Johnson (2023) argue that the advantages accruing to firms with the resources to develop and deploy frontier AI will produce unprecedented concentration of economic power rather than dispersal. Both predictions may be partially correct, and neither is separable from the material constraints that govern how quickly, and for whom, the agent economy can be built.
Access Is Not the Same as Capability
There is a challenge to the concentration argument that deserves direct engagement, because it is one that many readers will reasonably raise. AI agents are not confined to large organisations. Consumer products from OpenAI, Anthropic and others have made capable agents available to individuals, freelancers, and small businesses at relatively modest cost. Claude, ChatGPT, and comparable tools represent a genuinely meaningful democratisation of AI capability. If agents are already widely accessible, does the concern about concentrated access not lose its force?
The answer lies in a distinction that the binary framing of access versus no-access obscures. What matters is not whether organisations can access agents, but which tier of agent capability they can access. The agents available to an individual via a consumer subscription are genuinely useful for a wide range of discrete tasks. They can draft documents, summarise research, assist with analysis, and support decision-making in ways that would have been remarkable only a few years ago. However, they are not the same as the agents that a well-capitalised organisation can deploy when it builds proprietary systems on top of frontier models, fine-tuned on proprietary data, integrated deeply into operational workflows, and running at enterprise scale with priority compute allocation.
The difference is not merely one of degree. It is qualitative. A freelancer using a consumer agent can draft a contract more efficiently than before. A large firm deploying an enterprise agent can autonomously manage an entire procurement function across thousands of suppliers in real time, learning continuously from the data generated by each transaction. Both activities involve AI agents. They do not involve the same thing. The Coasean transformation, the genuine redrawing of firm boundaries, belongs to the latter category. It requires agents operating at a level of autonomy, reliability, and integration that consumer-tier access does not currently provide.
Shahidi et al. (2025) are instructive on this point. They observe that agent quality and therefore agent pricing is likely to scale with the compute allocated to it, with prices potentially rising in proportion to the stakes of the transaction. This means the most consequential deployments, those capable of substituting for teams of specialists rather than assisting individual workers, are anchored to a cost structure that reflects the infrastructure required to run them. Consumer access to agents is real and valuable. It does not, for the present, close the capability gap.
The Compute Imperative
The capability gap is, in significant part, a compute gap. The frontier AI models that underpin genuinely autonomous agents are among the most computationally intensive systems ever built. Training a frontier model requires data centre infrastructure consuming hundreds of megawatts of power and billions of dollars of capital expenditure (Sevilla et al., 2022). Inference, the process of running a model to complete a task, is cheaper per query but scales rapidly with deployment volume. The prospect of agents executing autonomous tasks at organisational scale implies a demand for compute and energy that current infrastructure is not positioned to satisfy without material investment and extended lead times.
The International Energy Agency estimated in 2024 that data centre electricity consumption could double by 2026, driven substantially by AI workloads (IEA, 2024). Goldman Sachs Research (2024) has estimated that each AI-assisted query consumes approximately ten times the electricity of a conventional web search, a differential that compounds rapidly at scale. The implication is that the Coasean singularity is not simply a function of algorithmic capability. It is also a function of whether the underlying physical infrastructure can support the scale of deployment that genuine organisational transformation requires.
This constraint creates a structural asymmetry that reinforces the capability tier argument. Shahidi et al. (2025) note that the economics of agent supply differ sharply between firms that train their own foundation models, incurring high fixed costs, and those that consume others' models as a service. The former category is currently limited to a small number of organisations with access to the capital, talent, and infrastructure required. This concentration at the foundation model layer propagates upward into the agent economy, and those with privileged access to compute have privileged access to the highest tiers of capability.
A Transitional Constraint with Lasting Distributional Consequences
It would be wrong to treat the compute constraint as permanent. The history of computing is a history of progressive dematerialisation. Recent developments, including the emergence of more efficient model architectures and purpose-built inference hardware, suggest that frontier-level performance may increasingly be achievable at lower cost (DeepSeek AI, 2025). Shahidi et al. (2025) anticipate this trajectory, noting that as returns to additional compute diminish, agents may eventually be offered at low cost or subsidised by complementary revenue streams, in patterns familiar from search and social media.
The transition will nonetheless take time, measured in years rather than months, a function of infrastructure investment cycles and semiconductor production timelines that software progress alone cannot compress. During this interval, the Coasean reorganisation of the firm will proceed unevenly. It will advance fastest in sectors with the capital and technical capacity to deploy at scale, and most slowly in those without. The concern is not that consumer access to agents will be withdrawn. It is that the agents capable of genuinely transforming organisational structure will remain, for a significant period, beyond the reach of the very organisations that might benefit most from them.
Institutional Implications
The twentieth-century firm was not only an economic institution. It was also a social one, providing employment, structure, identity, and community, and through the relationship between firms and collective bargaining, a mechanism for distributing economic gains between capital and labour. As AI dissolves the boundaries of the firm, these social functions must find new institutional homes. If they do not, the efficiency gains from AI adoption will be accompanied by social disruption that undermines the very stability those gains depend on.
This is not a new dynamic. The industrial revolution concentrated labour in factories, destroying the social fabric of rural communities and generating the political pressures that eventually produced the welfare state. The digital revolution concentrated economic power in platform companies, destroying the business models of industries from retail to journalism. The AI transition is likely to be faster and more pervasive than either predecessor. The institutions needed to manage its consequences, reformed labour market arrangements, new mechanisms for distributing the gains from AI adoption, redesigned education and training systems, need to be designed before the disruption becomes acute. Coase's insight was that institutional boundaries are not fixed. AI is changing the cost structure that determines them. The question is whether the social and political institutions that govern the economy can adapt with sufficient speed and intelligence to ensure that the resulting reorganisation produces broadly shared flourishing rather than concentrated wealth and distributed precarity. That is a question of institutional design. And institutional design, unlike technological development, is something that societies choose to do, or choose not to.
References
Acemoglu, D. and Johnson, S. (2023) Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity. New York: PublicAffairs.
Coase, R.H. (1937) 'The Nature of the Firm', Economica, 4(16), pp. 386-405.
DeepSeek AI (2025) DeepSeek R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. Technical Report. Available at: https://arxiv.org/abs/2501.12948 (Accessed: March 2025).
Goldman Sachs Research (2024) AI's Increasing Footprint in the Global Energy Sector. New York: Goldman Sachs.
International Energy Agency (2024) Electricity 2024: Analysis and Forecast to 2026. Paris: IEA.
Sevilla, J., Heim, L., Ho, A., Besiroglu, T., Hobbhahn, M. and Villalobos, P. (2022) 'Compute Trends Across Three Eras of Machine Learning', Proceedings of the International Joint Conference on Neural Networks. Piscataway: IEEE.
Shahidi, P., Rusak, G., Manning, B.S., Fradkin, A. and Horton, J.J. (2025) 'The Coasean Singularity? Demand, Supply, and Market Design with AI Agents', Working Paper. MIT and Harvard University.
Part of the Blueprint AI series, this post reflects my personal views and is independent of my professional role.