Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer

Abstract Consumers increasingly delegate purchasing decisions to Large Language Models (LLMs) acting as surrogate consumers. We develop “Tool-Lab,” an adaptation of information-board process tracing that places product attributes behind costly tool calls, and use it to examine how marketing pricing cues (i.e., just-below pricing and promotional framing) influence AI shopping agents. Across eight commercially deployed LLMs from three providers, we trace pre-choice information acquisition. Under zero cost, pricing cues rarely mislead. Imposing acquisition costs under a vague goal prompt leads LLMs to omit diagnostic attributes required to compute unit price and make suboptimal choices resembling human heuristics. Relative to a specific goal prompt that largely preserves diagnostic search and choice optimality, a vague goal prompt under constraints creates a search-mediated vulnerability for AI agents. This research demonstrates that the efficacy of marketing cues in delegated AI shopping is governed by storefront information architecture combined with consumers’ instructions, not necessarily immutable LLM characteristics.

Introduction. The evaluation of market alternatives has historically relied on human judgment. Increasingly, however, consumers are delegating end-to-end purchasing tasks to autonomous Large Language Models (LLMs) deployed as surrogate consumers (Pandya; Aw et al., 2025; Hasselwander et al., 2026; Puntoni and Wertenbroch, 2024), that is, an external entity to whom a principal relinquishes decision authority to navigate the marketplace (Solomon, 1986). Major platforms have begun deploying agents that search, compare, and complete purchases on consumers’ behalf (Verdon; Sataua; Kahn). While surrogate consumers have traditionally been humans, delegating this role to AI agents represents a fundamental shift in who evaluates the market (Dholakia et al., 2025).

Since for decades, retail pricing strategies, such as just-below pricing (Lacetera et al., 2012; Stiving and Winer, 1997; Thomas and Morwitz, 2005) and promotional discount framing (Inman et al., 1990; Thaler, 1985), have been engineered to exploit human bounded rationality, practitioners already advise retailers to structure pricing and product attributes so that AI agents can evaluate them (Cooper III). Still, when the buyer is an autonomous AI agent, it remains an open question whether pricing cues built for human cognitive limitations continue to influence purchasing, and if so, how.

Research evaluating LLMs as economic actors has found them to produce both rational (Brand et al., 2023; Cillo and Rubera, 2025; Horton et al., 2023) and heuristic-like behavior (Binz and Schulz, 2023; Filandrianos et al., 2025; Hermann and Puntoni, 2025). Reconciling this duality has been proven difficult because the existing literature predominantly treats heuristic-like behavior as an inherent, static defect hardwired into LLMs during training. As a result, research has largely documented whether LLMs exhibit heuristic-like behavior, but not what produces it.

We propose that heuristic-like purchasing in delegated AI shopping can be a property of the information state that the AI agent holds at the moment it commits to a choice. This pre-choice information state can be jointly shaped by retailers and consumers. Retailers determine the cost of retrieving product attributes through platform design, and consumers determine, through the specificity of their delegating prompt, which attributes the AI agent treats as diagnostic. The combination of costly retrieval and vague goal prompts leads AI agents to truncate their information search and choose suboptimally.

Related work. Marketing scholarship has examined Generative Artificial Intelligence (GenAI) primarily through firm-side strategy (Prasanna and Kushwaha, 2025; Riandhi et al., 2025; Grewal et al., 2025; Kumar et al., 2026), yet the critical frontier in marketing theory centers on the autonomous AI agent itself (Aw et al., 2025; Dholakia et al., 2025; Puntoni and Wertenbroch, 2024). Behavioral research in this domain predominantly treats Large Language Models (LLMs) as passive “silicon samples” or synthetic respondent pools (Joerling, 2025; Sarstedt et al., 2024) to emulate human survey responses and aid in marketing research (Arora et al., 2025; Li et al., 2025; Liu and Chen, 2026; Wang et al., 2026).

However, there is a growing use of AI agents by consumers as autonomous shopping assistants that perform end-to-end shopping (Hasselwander et al., 2026; Kim et al., 2023; Pandya; Google DeepMind, 2026). In marketing theory, delegating purchase authority to an external entity is a classic case of surrogate consumption (Solomon, 1986). A surrogate consumer is defined as an agent to whom a consumer relinquishes decision authority to evaluate, select, and acquire goods within the marketplace (Solomon, 1986). Today, generative AI, in particular Large Language Models (LLMs), can act as surrogate consumers because they are capable of autonomously querying APIs, exploring product catalogs, and executing financial transactions on a consumer’s behalf (Google DeepMind, 2026; Schick et al., 2023; Wang et al., 2024). In fact, Meta, Google, and Amazon now deploy LLM agents that complete purchases on consumers’ behalf (Meta; Sataua; Srinivasan; Amazon Staff), and nearly a quarter of U.S. consumers report making an AI-assisted purchase (Morgan Stanley).

As LLMs assume the role of surrogate consumers, a behavioral paradox emerges in the literature. LLMs neither conform strictly to classical economic assumptions of rationality (Thaler, 2000), nor do they predictably replicate human bounded rationality (Kahneman, 2003). On the one hand, an extensive body of research demonstrates that LLMs inherently possess strong analytical capabilities (Lee et al., 2026; Pandey and Singh, 2025). When prompted in standard discrete-choice environments, LLMs can function as highly capable simulated economic actors (Horton et al., 2023). They can evaluate expected values, respect downwardsloping demand curves, maintain budgetary compliance, and execute complex multi-attribute compensatory evaluations (Brand et al., 2023; Cillo and Rubera, 2025; Reisenbichler et al., 2026; Zhou et al., 2026). These competencies embody analytical (“thinking”) AI, characterized by computational precision, logic, and rule-based data processing (Huang and Rust, 2021).

Method. In economic decision-making, an agent evaluates market alternatives against a specific objective function. Depending on the decision context, this objective would optimize nominal expenditure, brand equity, perceived quality, or price per unit of volume. In the consumption of packaged goods, marketing theory and consumer economics establish that the normative objective is unit-price minimization—selecting the alternative j that minimizes financial expenditure per unit of physical quantity (Miyazaki et al., 2000; Russo, 1977):

Because computing this objective requires multi-step arithmetic, decision-makers operating under constraints frequently substitute the normative objective with a computationally simpler one, such as maximizing the discount percentage or minimizing the discounted list price. In marketing theory, this objective substitution is called transaction utility (Inman et al., 1990; Thaler, 1985).

Following this framing, we adopt unit-price minimization as the normative objective for consumer packaged goods and classify any choice that deviates from this normative objective as suboptimal. Heuristic then becomes a resource-conserving departure from full evaluation of the normative objective, either by substituting a simpler objective (e.g., price minimization) or by evaluating the normative objective on a reduced set of attributes (e.g., ignoring cents; Shah and Oppenheimer, 2008).

Product attributes can be classified by their role in the normative objective calculation. We refer to the subset of attributes that enter the unit-price equation as diagnostic (i.e., list price, whole dollars, cents, discount percentages, and weight), and the attributes that do not enter the equation as non-diagnostic (e.g., packaging, country of origin, calories).

In human judgment, information acquisition is costly. Human decision-makers face both external constraints (e.g., time pressure, fragmented shelf displays) and internal cognitive limitations (e.g., working memory capacity, computational fatigue; Payne et al., 1993; Shugan, 1980).

Similarly, an LLM operating as a surrogate consumer incurs both external and internal acquisition costs. External acquisition costs arise from storefront architecture and environment design, including the structural placement and prominence of attributes within the prompt or the structure of API payloads. Internal acquisition costs reflect computational resource allocation and LLM tendencies and capability. These include the LLM’s configured reasoning effort, the number of reasoning tokens it produces, or the number of tool calls it requests before producing the final response. LLM providers calibrate these propensities during fine-tuning to balance inference speed and API operating costs against output accuracy. If an AI agent were to exhaustively inspect every product attribute via unbounded tool calls, task execution would become prohibitively slow and expensive. Because online storefronts offer thousands of alternatives, each with many attributes, exhaustive inspection is prohibitively slow and would exceed the LLM’s output token limit. Thus, the LLM realistically must truncate its acquisition before inspecting every alternative and attribute combination. Throughout this research, we use the term acquisition cost to refer to both the internal and external constraints that an LLM operates within.

LLMs are instruction-tuned to deliver contextually coherent and relevant responses to users’ prompts. In a delegation relationship, however, task performance strongly depends on the goal specificity communicated by the delegating principal (Kleingeld et al., 2011; Locke and Latham, 2002). According to goal-setting theory (Latham, 2023), high goal specificity directs the agent’s effort to goal-relevant information.

We define prompt specificity as whether the prompt specifies the objective to be optimized. A specific goal prompt explicitly defines the objective function to be optimized (e.g., “Find the coffee with the lowest price per ounce”), and the diagnostic attribute set.

Discussion. This research examined whether pricing cues designed for human bounded rationality continue to influence AI shopping agents, and what governs when they do. Across eight commercially deployed LLMs from three major providers, just-below pricing and promotional framing affected delegated choices mainly through the information the agent held at the moment of choice, rather than through how the agent weighted the cues themselves. When attribute acquisition was costly and the delegating goal was vague, most evaluated LLMs omitted diagnostic attributes, and the resulting choices resembled human heuristic decisions.

When all attributes were visible, pricing cues rarely distorted choice. Seven of the eight LLMs weighted cents at the same rate as whole dollars (Table W2), and six of the eight LLMs weighted a dollar saved through a discount like a dollar spent through list price (Table W16). When attributes were hidden behind costly tool calls, the vague goal prompt negatively moderated the effect of acquisition cost on diagnostic acquisition, which in turn reduced choice optimality. The validation studies show that this degradation happens primarily through information acquisition rather than utilization. When information states were assigned exogenously, seven of the eight LLMs selected the alternative implied by the normative policy in 99.3% of sessions with partial information (Web Appendix F). When required to report unit values, no LLM made an arithmetic error (Web Appendix G). Selecting the lower whole-dollar or lower list-price alternative was therefore consistent with the information the agent held (Table W33, Table W34), even though it was economically inferior given the full attribute set.

Tool-Lab reveals which attributes were omitted. Under the specific goal prompt, acquisition cost primarily reduced acquisition of cents in Study 1 and discount percentage in Study 2, which are the same attributes the normative policy abandons first as cost rises (Web Appendix H). Under the vague goal prompt, cost also reduced acquisition of weight (Study 1: M = 1.10 of 2; Study 2: M = 1.31 of 3), which is a diagnostic attribute under a unit-price objective but non-diagnostic under a price objective. When price was consolidated into a single attribute, truncation shifted to weight, and the pathway replicated for the three Gemini LLMs tested (Table W11).

Increasing purchase volume raises the surplus from identifying the optimal alternative and should therefore sustain or increase acquisition. Instead, as we exogenously increased volume from one to 250 bags, Pro substituted the unit-price with a price minimization objective. The share of Pro’s pre-search reasoning traces that articulated a price objective increased from 37% to 86%, diagnostic acquisition fell from 5.55 to 4.22, and choice optimality fell from .80 to .35 (Figure W1). Holding volume constant, sessions articulating a price objective acquired fewer diagnostic attributes and, at 1, 5, and 50 bags, were 40 to 69 percentage points less likely to select the optimal alternative (Table W14). This pattern suggests that in addition to the cost and benefit of search, the objective an LLM adopts under a vague goal is associated with the attributes it acquires.

The hypothesized pathway generalized across most of the LLMs and providers we tested. All three hypotheses were supported for seven of the eight LLMs in Study 1 and six of the eight in Study 2, across the three providers (Table W6; Table W20). The pathway also replicated across most prompt phrasings and when price was consolidated into a single attribute (Table W8; Table W11). We found boundary conditions for two LLMs.

Limitations. Although the studies include robustness tests and validation experiments, several limitations qualify our findings and suggest directions for future research.

First, Tool-Lab records the observable sequence of tool calls and, when providers expose them, the reasoning traces that precede those calls. Like concurrent think-aloud protocols in human decision research (Ericsson and Simon, 1998; Payne et al., 1993), reasoning traces are verbal reports of how an LLM describes its task, not direct measures of the neural network activations that produce its behavior. Therefore, Tool-Lab shows what information an agent acquired and which objective it articulated, but not the internal mechanism that generated acquisitions and choice. Also, our results establish when search truncation produces heuristic-like choices, but they do not rule out rival accounts of why truncation occurs, such as statistical regularities in training data or limits on planning. In addition, reasoning traces were available only for LLMs whose providers expose them, and we classified traces only in the purchase volume study. Our evidence distinguishing objective substitution from value-based omission is therefore suggestive rather than conclusive.