The Coasean Singularity? Demand, Supply, and Market Design with AI Agents

Source: Peyman Shahidi, Gili Rusak, Benjamin S. Manning, Andrey Fradkin and John J. Horton; NBER Working Paper 34468 (chapter in The Economics of Transformative AI) · 2025-11

AI agents are autonomous software systems that perceive, reason, and act in digital environments to achieve goals on behalf of human principals, with capabilities for tool use, economic transactions, and strategic interaction. This chapter focuses on the transformative implications of these systems as market participants. We envision agents as having the ability to harness computational resources, to communicate with other agents and humans, to receive and send money, and to access and interact with the Internet. A variety of agents are already available (e.g., AutoGPT, Replika, Claude Code), and more capable ones are in the development pipeline.

We offer a practical perspective on AI agents as participants in digital markets, complementing Hadfield and Koh’s (2025) more theoretical treatment in this volume. We take seriously the idea that, rather than hiring a human agent, one could instead rely on an AI agent. This possibility is near, as agents already perform, albeit imperfectly, tasks such as shopping (Dammu et al. 2025), negotiation (Zhu et al. 2025), or search for products on e-commerce sites (Zeff 2025; Allouah et al. 2025). It is easy to imagine more advanced agents contacting multiple counterparties to negotiate prices or applying to jobs and advocating for employment on a user’s behalf.

The fundamental economic promise of AI agents lies in their ability to dramatically reduce transaction costs—the expenses associated with using markets to coordinate economic activity. This reduction occurs not only through direct task execution but also via advisory services. In the direct approach, agents perform tasks entirely on behalf of users, from price comparison shopping to contract negotiation to job interviews. In the advisory approach, agents assist users in making better market decisions, such as helping job applicants individually optimize their resumes for each job submission. Recent evidence suggests both approaches can be quite effective. For the former, at least one study has shown that outcomes can be better for job seekers interviewed by AI as opposed to human employers (Jabarian and Henkel 2025). For the latter, algorithmic assistance on job application essays can increase hiring rates significantly (Wiles et al. 2025).

It is important to recognize that demand for AI agents represents derived demand rather than direct utility. Individuals do not generally derive satisfaction from watching an agent compile price lists for gas grills; they employ the agent purely to achieve some desired market outcome following their own decision-making process. As a result, humans will deploy AI agents in two primary scenarios: first, to optimize decisions that they would otherwise make sub-optimally due to information constraints or cognitive limitations; and second, to make decisions of similar or potentially even lower quality, but at dramatically reduced cost and effort.

The broad adoption of AI agents will cause transformative downstream effects on the economy. Although the exact nature of these changes remains uncertain, the forces at play are familiar to economists: supply and demand of AI agents will continue to shape market organization, while technological change alters the relative costs of different activities. To see how these forces might operate, Coase’s (1937) insight that transaction costs play a central role in shaping organizations is particularly useful. One could argue that much of how we structure our economy and firms can be explained by transaction costs, often costs of human labor. The activities that comprise transaction costs—learning prices, negotiating terms, writing contracts, and monitoring compliance—are precisely the types of tasks that AI agents can potentially perform at very low marginal cost. Once agents can indeed execute these functions effectively and cheaply, we will see significant shifts in the traditional make-or-buy boundaries that define firm organization and market structure.

Humans will demand AI agents for the same reasons they demand human agents: they believe it is rational to have some aspect of a decision or market transaction handled through an intermediary. This might occur because the agent’s time is less expensive than the principal’s time, because the agent is legitimately better at the task than the principal, or because the principal has reasons for obscuring their identity in the transaction.

These motivations will increasingly drive the adoption of AI agents along two avenues. First, agents will substitute for human intermediaries where one would otherwise do the work personally or hire a human agent. The canonical example is product search. AI agents convert the costly, time-consuming parts of intermediation—search, screening, quoting, negotiation, scheduling—into low-cost compute and API calls. An agent can solicit and compare many quotes in parallel, then book and monitor follow-through at far lower marginal cost than a human. Even when skilled execution remains human, the intermediation premium falls, creating demand for AI agents.

Second, and perhaps more consequential, agents will enable undertakings that would not have been attempted at all. By lowering the cost of exploration and execution, they expand the feasible set of options and reduce the threshold of “worth doing.” For physical jobs (e.g., home installation, fixing a sink), the agent can conduct diagnostic triage, source parts, select vendors, and schedule service. For software, it can generate and iterate on a bespoke script. Compared to humans, agents can persist through repeated failures at much lower marginal cost and continue monitoring tasks over longer windows to secure better outcomes.

AI agents are likely to first gain traction in markets where human agency is already common. Key characteristics of these markets include high-stakes transactions, large pools of potential counterparties, substantial effort required to evaluate options, information asymmetries that could be resolved through effort or experience, and experience asymmetries due to differences in transaction frequency. Examples of such markets include job search, real estate, and certain investment contexts where counterparties matter beyond financial terms alone. A useful heuristic suggests AI agents will appear first in markets that already employ human agents or where large digital platforms were created to overcome matching frictions—such as LinkedIn, Upwork, Zillow, and Airbnb. Table 1 summarizes these characteristics and associated markets, highlights the existing human or platform-based solutions that address them today, and shows how AI agents can improve on those approaches. As adoption spreads from these initial footholds, the dynamics of trust, evaluation, and inter-agent coordination will shape broader economic integration.

Another equally important dimension of supply is agent pricing. Human agent services are often priced as a percentage of the transaction value. The sums involved can be high in situations where the human agent is paid to engage in an adversarial situation against another human agent, and where one side wins and another loses. These high sums in adversarial settings are due to two types of market forces: that the “best” human agents are scarce and that high-powered incentives result in extraordinary effort.

AI agents, however, face a very different cost and supply profile, which changes the pricing dynamics. Because software can be copied at negligible cost, the supply of AI agents will be virtually unconstrained. Unlike human agents, they also do not derive utility from pecuniary compensation. These factors point to diminished pricing power for AI relative to human agents. However, an opposing force pushing toward higher agent prices arises if the quality of an agent’s performance depends on the compute allocated to it. In that case, prices could scale with the amount of compute required—potentially rising in proportion to the stakes of the transaction.

Orthogonal to ownership is specialization. We posit that there are likely to be two types of AI agents along this dimension: horizontal and vertical. Horizontal agents are generalists that can span many tasks and platforms, leveraging a single memory/preferences layer across markets and arbitraging options when beneficial, but they may lack domain-specific tooling or compliance features. Vertical agents, on the other hand, specialize in a narrow domain (e.g., tax filing, job search, travel) or even a specific platform workflow, trading breadth for depth (i.e., stronger performance, richer integrations, and tailored guardrails) at the cost of portability and reuse.

Perhaps the best cases for AI agents reducing economic rents are in markets where firms exploit behavioral biases and bounded rationality. For example, many consumers choose suboptimal contracts given their expected usage patterns, leading to rents for firms (e.g., see Grubb (2009) on phone contracts). By making more rational decisions on behalf of consumers, agents would render such rent-extraction strategies less viable, pushing markets closer to the competitive ideal.

AI agents will also affect the prevalence and dynamics of bargaining as their opportunity cost is not human time. Classic bargaining models assume impatience and a preference to conclude negotiations quickly because time and attention are scarce resources for humans. However, for the AI agent the binding constraint is compute rather than time. As a result, even if the principal prefers timely resolution, agents can initiate negotiations earlier, hold many concurrent negotiations, and continue engaging in these for a longer period of time (e.g., begin negotiating summer-2027 rentals in January 2026).

Beyond bargaining dynamics, AI agents will also alter the prices of goods and services, as agents’ ability to elicit and reveal preferences enables new mechanisms. When consumers willingly and accurately reveal their preferences to trusted agents, firms gain access to richer, more detailed preference information. This enhanced informational access enables personalized pricing strategies that can improve price discovery and market efficiency, lowering the deadweight loss previously caused by information asymmetry. However, efficiency improvements do not determine the distribution of rents. While personalized pricing may benefit consumers by closely matching products to individual preferences, it may also lead to inequitable outcomes if firms exploit consumer-specific price elasticities. At the same time, consumers’ agents may strategically withhold certain information to secure more favorable prices, adding additional complexity to the market.

A fundamental challenge of implementing new market designs emerges from the reality that most Internet activity will eventually originate from AI agents rather than humans. These agents will be able to effectively mimic humans in many situations. For example, advertisements are priced based on impressions and clicks, but agents will also be able to “see” and click on them. Similarly, social media companies and their users would like to clearly verify that content is produced by a specific human rather than a bot. Sybil attacks—where single entities create multiple fake identities to gain disproportionate network control—become more difficult to detect in agent-mediated environments. This necessitates new approaches to digital accountability, potentially including cryptographic proofs, digital IDs, or decentralized verification platforms to ensure system safety and reliability.

Of particular importance is identity verification, both for humans and for AI systems that are acting on behalf of specific humans. Two broad classes of solutions emerge: walled-garden approaches where platform gatekeepers exclude potentially malicious actors, and open systems with robust verification mechanisms.