Agentic markets

Source: Martin Bichler; Electronic Markets (Springer), open access · 2026-06-06

We refer to such settings as agentic markets: electronic markets in which autonomous AI agents act on behalf of buyers and/or sellers to execute search, evaluation, negotiation, and transaction decisions. In this paper, a market is understood as a decentralized electronic exchange environment in which prices, search, and matching determine trade outcomes. A trade denotes the matching and transaction outcome resulting from these search and pricing processes.

The central question we address is: How does delegation to AI agents alter equilibrium prices, entry, and matching efficiency in markets characterized by search frictions and product differentiation?

We argue that agentic markets do not eliminate market frictions; they transform them. While AI agents can drastically reduce technical search costs, delegation to probabilistic, content-generating systems introduces new behavioral and informational frictions. Rather than converging toward frictionless competitive equilibria, agentic markets may sustain price dispersion, inefficient entry, and matching distortions through transformed economic primitives. These arguments challenge the common presumption that AI-mediated markets necessarily approximate Bertrand-style competition as search costs approach zero.

We identify five capabilities of GenAI-based agents that are economically significant for electronic markets. These capabilities modify core primitives of standard microeconomic models and therefore shape equilibrium outcomes in agentic markets.

First, autonomous delegated search allows agents to interpret loosely specified natural-language objectives. Rather than operating on fully specified utility functions, AI agents infer latent preferences, which might also introduce potential noise in match evaluation. This replaces explicit preference revelation with probabilistic inference.

Second, they can inspect large numbers of alternative offerings. This reduces technical search costs and shifts the locus of search from human cognition to algorithmic mediation.

Third, AI agents can summarize and compare product information. While this can reduce information overload, it may also compress or distort match-value signals, affecting effective product differentiation.

Fourth, Automated negotiation and execution allow agents to initiate transactions, trigger renegotiations, and respond dynamically to price changes. This creates the possibility of rapid strategic interaction and continuous price adaptation.

Fifth, low-cost content and offer generation reduces barriers to entry by lowering the cost of creating listings, product variants, or digital goods. This can increase variety but may also induce excessive entry and low-value proliferation.

In summary, with vastly decreasing search costs in agentic markets, competition intensifies because consumers can easily check “one more option” to find a better deal. This intensifies competitive pressure and can reduce markups. Consumers search more exhaustively (reservation utility rises), ensuring they find the product that best fits their specific preferences rather than settling for a “good enough” option. Although there is some tendency toward excessive entry even at low search costs, in a low-friction world, the model predicts an efficient market.

However, a growing body of experimental and simulation-based evidence suggests that LLM-based agents do not always search in a fully frictionless or fully rational manner. Across a range of environments, some agents exhibit model-dependent order effects, limited exploration, early stopping, and sensitivity to prompt framing (Bansal et al., 2025). While the direction and magnitude of these behaviors vary across models and decision settings, they indicate that even when the technological cost of inspecting additional offers is negligible, residual behavioral frictions may persist and affect search depth and matching outcomes.

Despite the potential for automation and efficiency gains, experimental evidence indicates that AI agents do not always implement fully rational search strategies. Importantly, the documented biases are not uniform across models. Allouah et al. (2025) show that ranking effects depend on the underlying LLM architecture: some agents display stronger positional preferences than others, and these preferences need not take the form of a simple bias toward the first displayed option. Similarly, Bansal et al. (2025) report that some agents exhibit order effects or degraded performance when faced with larger option sets, while the direction and magnitude of these effects vary across models and decision stages.

The aggregate market effect depends on the composition of agent types. If many agents in a market rely on similar architectures and exhibit systematically limited exploration, the resulting reduction in effective search can weaken competitive pressure. In more heterogeneous environments, different positional preferences may partially offset each other. However, even in such cases, market-wide exploration may remain incomplete if multiple agent types sample only a limited subset of the available alternatives.