AI shopping agents lower technical search costs but do not guarantee better markets; quirks in agent behavior, flood of AI-generated offers and easy entry can weaken signals and leave digital marketplaces inefficient.
Agentic markets
Generative-AI agents can lower technical search costs but behavioral departures from optimal search, AI-driven signal dilution, and low-quality entry can create friction that traps markets in inefficient equilibria.
Abstract Generative AI enables autonomous software agents that can search, compare, and transact across digital marketplaces, promising large reductions in consumer search costs and improved matching between buyers and sellers. This paper argues that such gains are not automatic. Drawing on economic search theory, we first discuss the impact of reduced search costs on markets. Then, we show how the behavior of current AI agents introduces frictions that limit competitive outcomes. Empirical studies reveal persistent deviations of AI agents from optimal search behavior that function as behavioral search costs even when technical search costs approach zero. At the same time, AI-generated content contributes to signal dilution, reducing the informativeness of offers by AI agents and weakening effective product differentiation. These forces interact with low entry costs, which encourage excessive and often low-value entry. Together, they can trap agentic markets in inefficient equilibria. We outline key implications for electronic marketplaces and highlight promising directions for future research on agentic markets.
Summary
Main Finding
Generative-AI–mediated “agentic markets” do not automatically converge to frictionless, Bertrand-style competition as technical search costs fall. Instead, probabilistic, content‑generating AI agents transform search frictions and information primitives so that behavioral and informational frictions (e.g., order effects, limited exploration, hallucinations, signal dilution), together with much lower entry costs, can sustain price dispersion, induce excessive low‑value entry, compress effective differentiation, and trap markets in inefficient equilibria.
Key Points
- Definition: Agentic markets are decentralized electronic markets where autonomous LLM‑based agents act for buyers and/or sellers to search, evaluate, negotiate, and transact.
- Five agent capabilities with economic consequences:
- Preference interpretation from loosely specified natural language → introduces preference‑inference noise (constructs utility probabilistically).
- Autonomous delegated search across platforms → reduces technical search costs but can leave residual exploration frictions (order effects, early stopping), producing persistence of price dispersion.
- Summarization and comparison of product information → can compress or distort match signals, causing signal dilution and weaker effective differentiation.
- Automated negotiation and execution → enables rapid price dynamics but introduces strategic and reliability noise (stochastic pricing deviations, unstable bargaining).
- Low‑cost content/offer generation → dramatically lowers entry costs, increasing listing volume and encouraging excessive, often low‑value entry.
- Mapping to economic primitives: agent behaviors change search cost (c), signal precision / differentiation intensity (μ), and entry costs; these shifts are often endogenous to platform and system design (ranking, interfaces, verification, data portability).
- Theoretical lens: uses the Anderson & Renault (1999) search-and-differentiation framework (and contrasts with Bertrand, Diamond, Chamberlin) to show how transformed primitives determine equilibrium prices, search intensity, and entry.
- Empirical regularities motivating the argument: documented LLM‑agent behaviors such as position bias/order effects, limited exploration, hallucinations, signal compression (cited recent experimental/simulation studies, 2023–2025).
- Policy and design point: platform architecture and governance (ranking algorithms, interface constraints, verification, agent transparency) materially shape market equilibria in agentic markets.
Data & Methods
- Paper type: conceptual Foundations article (no original econometric dataset).
- Evidence base: structured literature search covering 2023–2025 across Google Scholar, SSRN, arXiv, AIS eLibrary, and leading AI/IS conference proceedings (AAAI, ICML, NeurIPS, ICIS), plus economics/IS journals.
- Inclusion criteria: empirical, experimental, and simulation studies of LLM‑agent behavior in market‑relevant tasks (search, ranking, pricing, negotiation, matching). Excluded purely technical benchmarks without market interaction.
- Analysis method: mechanism‑based coding of selected studies to extract recurring behavioral patterns (order effects, early stopping, hallucinations, signal compression, stochastic negotiation) and map them to economic primitives (search cost c, differentiation μ, entry cost).
- Theoretical framework: embeds these mapped changes into the Anderson & Renault (1999) model of costly sequential search with differentiated products and endogenous entry to derive comparative static and equilibrium implications.
- Limitations: not an exhaustive systematic review; aims to synthesize recurring regularities into a coherent microeconomic mapping rather than to provide new empirical estimates.
Implications for AI Economics
- Equilibrium outcomes
- Zero technical search cost does not guarantee Bertrand outcomes: behavioral search frictions can sustain search stopping and price dispersion (Diamond‑like or intermediate outcomes).
- Signal dilution from AI content weakens effective differentiation (reducing informativeness of offers), which can both reduce welfare (buyers get worse matches) and shift pricing in nontrivial ways (search vs. market‑power tradeoffs).
- Lower entry costs can lead to excessive entry and proliferation of low‑value offerings, exacerbating search burdens and lowering average match quality.
- Welfare and market structure
- Agentic markets can be trapped in inefficient equilibria: higher welfare losses than predicted by purely technical cost reductions because of agent behavioral biases and noisy signals.
- Increased platform gatekeeping: when agents rely on platform APIs, discovery/control shifts toward platforms (ranking and routing matter more), with implications for competition and market power.
- Platform and regulatory design levers
- Governance matters: platform choices (ranking algorithms, presentation/interface, verification/certification of agent outputs, audit/sandbox environments, limits on automated listing creation) directly affect c, μ, and entry incentives.
- Mitigation options include forcing or encouraging agent transparency (explainable recommendations), standardized signal/attribute formats to reduce signal dilution, verification/certification of agent‑generated listings, throttling mass automated entry, and audit trails to detect hallucinations or manipulation.
- Research directions for AI economics
- Empirical quantification: measure “behavioral search costs” induced by LLM agents (e.g., reservation thresholds, exploration intensity) and quantify signal‑dilution effects on match quality.
- Market experiments and simulations: multi‑agent lab/field experiments and agent‑based simulations to study dynamics with heterogeneous agents, strategic seller algorithms, and platform policies.
- Mechanism and platform design: design ranking, verification, and interface mechanisms that internalize agentic frictions and restore competitive pressure and efficient matching.
- Dynamic and strategic analysis: study long‑run entry dynamics, endogenous invention of product attributes by agents, and strategic interaction of agentic buyers and seller pricing algorithms.
- Policy evaluation: welfare analyses of disclosure, certification, and limits on automated content generation; antitrust implications when platforms act as gatekeepers for agentic discovery.
- Broad takeaway for economists and platform designers: when market actors are AI agents, economic analysis must treat technical capabilities, agent behavioral biases, and platform architecture jointly — platform governance and system design are economic variables, not only technical implementation details.
Assessment
Paper Typetheoretical
Evidence Strengthlow — The paper is primarily theoretical and conceptual; its empirical components are described as 'empirical studies' showing deviations but the abstract indicates descriptive/simulation evidence rather than causal identification, so claims about market outcomes rely on theory and observational/descriptive findings without strong causal inference.
Methods Rigormedium — Theoretical grounding in economic search theory appears careful and useful, and the paper also uses empirical studies and simulations to illustrate points; however, the abstract suggests these empirical elements are descriptive and not backed by robust identification strategies, representative datasets, or randomized interventions, limiting methodological rigor for causal claims.
SampleA mix of analytical economic models, simulations, and observational/descriptive empirical studies of current generative-AI agent behavior in digital marketplaces (details and sampling frame not provided in the abstract).
Themesadoption innovation
GeneralizabilityFindings hinge on current-generation generative models and agent designs and may not generalize as agent capabilities or incentives evolve, Results may differ across marketplace types (digital goods, services, physical goods) and platform architectures, Empirical illustrations appear to draw on specific settings or simulations, limiting external validity to real-world, large-scale markets, Regulatory, institutional, and geographic differences in platforms and market structure may alter outcomes
Claims (5)
| Claim | Direction | Confidence | Outcome | Details |
|---|---|---|---|---|
| Generative AI enables autonomous software agents that can search, compare, and transact across digital marketplaces, promising large reductions in consumer search costs and improved matching between buyers and sellers. Consumer Welfare | positive | high | consumer search costs and matching quality between buyers and sellers |
0.02
|
| Gains from autonomous AI agents (reduced search costs and improved matching) are not automatic because the behavior of current AI agents introduces frictions that limit competitive outcomes. Market Structure | negative | high | degree of competition / competitive outcomes in markets with AI agents |
0.12
|
| Empirical studies reveal persistent deviations of AI agents from optimal search behavior that function as behavioral search costs even when technical search costs approach zero. Task Completion Time | negative | medium | deviations from optimal search behavior (behavioral search costs) |
0.07
|
| AI-generated content contributes to signal dilution, reducing the informativeness of offers by AI agents and weakening effective product differentiation. Market Structure | negative | medium | informativeness of offers / effective product differentiation |
0.07
|
| Low entry costs into agentic markets encourage excessive and often low-value entry, and together with behavioral frictions and signal dilution can trap agentic markets in inefficient equilibria. Market Structure | negative | medium | market efficiency / equilibrium outcomes (inefficient equilibria) |
0.01
|