When Agents Shop for You: Role Coherence in AI-Mediated Markets

Abstract: Consumers are increasingly delegating purchase decisions to AI agents, providing natural-language descriptions of their preferences and identity. We argue that these representations constitute an information channel, role coherence, through which sellers can infer willingness to pay without explicit disclosure by the buyer agent, leading to preference leakage. In an experiment where a language-model buyer agent shops on behalf of a verbal consumer profile, we show that seller-side inference from dialogue alone recovers willingness to pay nearly one-for-one. Comparing this setting to a numeric-budget condition with confidentiality instructions cleanly isolates role coherence as distinct from instruction-following failure. Because this leakage arises from delegation itself, it cannot be mitigated at the prompt level. Instead, we propose architectural interventions that trade off personalization against preference privacy.

• The paper demonstrates that verbal persona instructions yield nearly one-for-one inference of consumer WTP, unlike numeric-budget directives which compress signals.
• It employs rigorous experimental designs with controlled trials (N=60 per cell) to contrast role coherence from traditional instruction-following mechanisms.
• Findings highlight privacy risks in AI-mediated commerce, urging the development of regulatory and technical countermeasures for personalized pricing.

Role Coherence as a Preference Leakage Channel in AI-Mediated Markets

Introduction

The increasing deployment of AI agents as intermediaries in consumer markets fundamentally alters information transmission dynamics between buyers and sellers. The paper "When Agents Shop for You: Role Coherence in AI-Mediated Markets" (2604.26220) rigorously analyzes how consumer delegation to AI agents—via natural-language persona descriptions—establishes a novel information channel termed "role coherence," enabling seller-side inference of private willingness to pay (WTP) even when explicit numeric budgets or financial disclosures are absent or forbidden.

Framework: Role Coherence versus Instruction-Following

The study distinguishes between two modes of delegation: verbal persona profiles and numeric budget instructions. In the verbal condition, a consumer’s description (e.g., occupation, lifestyle, preferences) is supplied to the buyer agent, which tailors shopping behavior accordingly. In the numeric condition, a budget value is provided but is accompanied by strict confidentiality directives. The core claim is that role coherence—behavioral fidelity to the supplied persona—is inherently correlated with WTP and transmits preference information through agentic behavior, independent from explicit numeric tokens or privacy violations.

The theoretical predictions are formalized using two assumptions:

• Privacy-directive invariance: Numeric-budget instructions render agent behavior invariant to the actual budget, compressing shopping signals toward the population mean.
• Character-driven generation: Verbal-instruction agents generate distinctive behavioral distributions contingent on the specific persona description, enabling rank-accurate and level-accurate inference of WTP.

Empirical contrast between the two channels isolates role coherence as a distributional leakage mechanism, distinct from instruction-following failure.

Experimental Design

A buyer-side AI agent (Claude Haiku 4.5) is prompted with either a numeric budget (with privacy directive) or a verbal persona. The agent interacts across multiple conversational turns with a seller agent offering a fixed five-product catalog. Each interaction transcript is then assessed by an inference agent tasked with estimating the buyer’s WTP based solely on dialogue content, without access to catalog or explicit price reference. The design ensures statistical rigor: $N=60$ trials per cell, six WTP levels (\$50–\$500), both full and redacted transcript variants.

A comprehensive factorial extension evaluates profile and prompt variance, crossing three persona variants and two prompt scaffolds per WTP cell, thus bounding design-induced variance.

Empirical Results

The central empirical finding is the slope of inferred WTP on target WTP across conditions.

(Figure 1)

Figure 1: Inferred versus target willingness to pay, by condition; verbal instructions yield a near one-for-one slope, numeric compress to the prior mean (slope 0.21), with non-overlapping confidence bands across all cells.

Verbal Profile Condition

• Accuracy: Aggregate MAE is \$48; 58% of trials within 25% of target WTP.
• Linearity: OLS slope of mean inferred WTP on target is 1.00 (95% CI [0.96, 1.05]); Spearman rank correlation 1.00.
• Robustness: Stripped-vocabulary profiles (no financial terms) retain strong signal (slope 0.85, CI [0.79, 0.91]); persona-redacted transcripts (demographic cues removed) preserve slope (0.93, CI [0.89, 0.97]).
• Factorial extension: Cross-profile and prompt variance minimally affect slope; the channel’s robustness is not contingent on specific profile or prompt wording.

Numeric Budget Condition

• Compression: Slope is 0.21 (CI [0.17, 0.26]); MAE \$92; only 36% accuracy within 25% of target.
• Behavioral Invariance: Estimates are prior-dominated and insensitive to true budget.

Redacted Transcript Analyses

• Without dollar anchoring, verbal estimates remain rank-correct but lose level calibration, inflating at higher tiers (slope 2.90). Numeric-redacted inference is unreliable, flattening near the population mean.

Mechanistic Implications and Robustness

Role coherence operates through behavioral generation: feature priorities, reactions to prices, product comparisons, and purchasing decisions. Persona cues are non-essential—the behavioral signature suffices for reliable inference. Stripped-vocabulary and persona-redacted conditions demonstrate that the channel persists in the absence of lexically explicit or demographic cues.

The numeric-budget paradigm’s failure to convey WTP through behavior emphasizes role coherence as fundamentally distinct from instruction-following failures reported in prior agentic negotiation literature.

Practical and Theoretical Implications

• Agentic Commerce: Seller-side inference of WTP via agentic behavior poses new privacy risks unmitigated by prompt-level privacy directives, as leakage is intrinsic to behavioral fidelity rather than token-level disclosure.
• Personalized Pricing: Role-coherent leakage introduces a powerful, non-cookie-based source of WTP information relevant for personalized pricing and price discrimination strategies.
• Policy and Platform Design: Regulatory interventions targeting explicit data fields are insufficient. Architectural solutions (e.g., anonymizing intermediaries, persona rotation, federated aggregation) are required for effective preference privacy.
• Future Directions: Extending to cross-model, cross-scenario replication will validate the generality of the channel. Real-world deployment analysis and countermeasure implementation are urgent next steps.

Conclusion

Role coherence is empirically validated as a powerful preference leakage channel in AI-mediated markets. Verbal persona delegation fundamentally generates behavioral signals from which seller-side agents infer WTP with high fidelity, irrespective of explicit numeric disclosures or privacy instructions. These findings demand a reconsideration of privacy architectures and regulatory frameworks in agentic commerce, as behavioral leakage arises from the very mechanisms that enable helpful delegation and personalization.