As large language models become autonomous consumer agents, standard consumer theory needs revision — agentic markets can break assumptions like direct revealed preference and individual rationality, creating new alignment and market-structure challenges.

Main Finding

The paper proposes "LLM Consumer Behavior Theory," a conceptual framework that reconceptualizes consumer theory for markets where large language model (LLM)–based agents make consumption decisions on behalf of humans. It maps how human preferences are represented, instantiated, and (mis)aligned in LLM agents, and how agent-level choices aggregate into market-level demand — highlighting important divergences from classical consumer models (notably risks of reduced heterogeneity, alignment failures, and shifting market dynamics). The contribution is synthetic and theoretical: it organizes prior NLP and economic work into a coherent research agenda rather than providing new empirical validation.

Key Points

• Motivation
  • Agentic AI (LLM agents) are increasingly making consumption choices for users; McKinsey and others forecast large-scale agentic commerce by 2030.
  • Classical consumer theory (humans as decision-makers) must be extended to handle delegated, agentic decision-making.
• Conceptual architecture
  • Three linked components:
  • Consumer preferences: grounded in classical utility and random utility theory, extended with behavioral economics (biases, bounded rationality).
  • User–agent alignment: principal–agent framing — agents infer proxies for user preferences via prompts, history, personas, or fine-tuning; alignment is imperfect.
  • Consumer markets: aggregation of agent-level willingness-to-pay leads to an "agentic market demand" that can differ qualitatively from human-driven demand.
• LLM decision properties
  • LLMs show a mix of rational-like behavior and systematic deviations (e.g., different time preferences, risk/loss attitudes, proneness to nudges).
  • Larger/recent models often give more stable preferences; choices can be highly prompt- and design-sensitive.
• Alignment mechanisms and limits
  • Persona assignment, few-shot/in-context learning, and fine-tuning can improve alignment but have limits (inconsistency, stereotyping, scalability/cost).
  • New approaches like low-rank reward (LoRe) attempt scalable personalization.
• Market-level implications
  • Risk of homogenization: LLMs tend to reproduce modal/average patterns from training data and share similar architectures and corpora, possibly reducing heterogeneity in aggregate demand.
  • Homogenization can concentrate demand on a few "modal" products and make market shares unstable following model updates.
  • Sources of agentic heterogeneity remain: model architecture, temperature/settings, and explicit personalization.
• Open research directions and ethics
  • Empirical questions: stability and robustness of LLM choice patterns; steering/controllability; extent of homogenization.
  • Personalization questions: what data/alignment best capture true preferences; minority representation; costs of individualized models.
  • Market dynamics: hybrid markets, simultaneous agent updates, equilibrium and pricing effects, accountability and consent, transparency and fairness.
• Limitations noted by authors
  • The paper is a theoretical synthesis (no new empirical validation).
  • Focus limited to demand side (not agentic supply or dual-agent markets).
  • Not an exhaustive literature review.

Data & Methods

• Methodological approach
  • Conceptual synthesis: integrates classical consumer theory (utility maximization, budget constraints, random utility), behavioral economics (biases), and recent NLP/agent literature.
  • Formal framing: applies principal–agent models to map user preferences → proxy representations → agent instantiation → agent-level choices → aggregated agentic demand.
  • Uses standard economic constructs (utility U(x) = V + ϵ, Willingness to Pay, demand curve aggregation) to interpret LLM behaviors.
• Evidence base surveyed
  • Empirical and experimental NLP/AI studies on LLM decision-making, preferences, risk and time attitudes, persona simulation, and personalization (examples cited: discrete choice/WTP studies, decision experiments, persona and fine-tuning research).
  • Studies documenting model sensitivity to prompts, updates, temperature, and alignment methods.
• What the paper does not do
  • No original data collection or empirical estimation.
  • No market equilibrium simulations or welfare calculations beyond conceptual discussion.

Implications for AI Economics

• Demand-side transformation
  • Agentic consumers change the mapping from individual preferences to market demand; economists must model the preference-reflection pipeline (user → agent proxy → agent choice).
  • Aggregation of agentic choices can yield different demand elasticity, concentration, and product share dynamics than human-driven markets.
• Heterogeneity and welfare
  • Reduced heterogeneity (homogenization) could lower demand diversity, harm niche products, and concentrate surplus toward a narrower set of offerings — raising distributional and welfare concerns for minority preferences.
  • Sudden model updates could cause abrupt market-share shifts, creating volatility and coordination externalities.
• Market structure and competition
  • Concentration of LLM providers and shared model weights may function similarly to platform effects, potentially amplifying market power and affecting pricing strategies and product differentiation.
  • Firms may strategically engineer offerings for modal agent preferences, changing marketing and product design incentives.
• Personalization and segmentation
  • Effective personalization (in-context learning, fine-tuning, low-rank reward methods) is central to preserving heterogeneity and fairness, but has cost, scalability, and privacy trade-offs.
  • The economics of personalization (when to share vs. individualize agents) becomes a key design and regulatory question.
• Policy, regulation, and accountability
  • New regulatory concerns: informed consent for delegated decisions, accountability for agentic mistakes, transparency of alignment processes, and protection of minority tastes.
  • Antitrust and consumer-protection agencies may need to consider agentic coordination risks (e.g., simultaneous updates causing market shocks).
• Research agenda for AI economists
  • Empirically measure LLM-induced changes to demand curves, elasticity, and consumer surplus.
  • Model hybrid markets where human and agentic decision-making coexist; analyze equilibria and dynamic responses to mass agent updates.
  • Study welfare impacts across heterogeneous consumers and distributional consequences of agentic homogenization.
  • Evaluate policy interventions (mandates for transparency, limits on shared model updates, standards for consent and accountability).

Summary note: the paper sets up a disciplined research program at the intersection of economics and NLP — providing conceptual tools and an agenda but leaving empirical validation and formal quantitative models for follow-up work.