arxiv: 2604.26220 · v1 · submitted 2026-04-29 · 💻 cs.MA · econ.GN· q-fin.EC

Recognition: unknown

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

Soogand Alavi , Salar Nozari

Authors on Pith no claims yet

Pith reviewed 2026-05-07 12:56 UTC · model grok-4.3

classification 💻 cs.MA econ.GNq-fin.EC

keywords AI agentsbuyer agentsrole coherencepreference leakagewillingness to payAI-mediated marketsdelegation

0 comments

The pith

Delegating purchases to AI agents with natural-language profiles lets sellers infer willingness to pay from dialogue alone.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Consumers increasingly let AI agents handle purchases by supplying verbal descriptions of their preferences and identity. This setup creates an information channel through which sellers can recover the consumer's willingness to pay nearly one-for-one from the resulting conversation. The paper isolates this effect by comparing it to a condition where agents receive numeric budgets plus explicit confidentiality instructions. The leakage persists because it stems from the delegation itself rather than from agents failing to follow rules. The authors conclude that prompt-level safeguards are ineffective and that higher-level architectural changes are required to manage the resulting privacy trade-off.

Core claim

When language-model buyer agents shop on behalf of verbal consumer profiles, seller-side inference from dialogue alone recovers willingness to pay nearly one-for-one. Comparing this outcome to a numeric-budget condition with confidentiality instructions shows that the leakage arises from role coherence rather than from instruction-following failure. Because the channel is created by the act of delegation to an agent using natural language, it cannot be closed at the prompt level. Architectural interventions are therefore needed to trade off personalization against preference privacy.

What carries the argument

Role coherence, the information channel through which sellers infer willingness to pay from natural-language buyer dialogues without explicit disclosure.

If this is right

Seller agents can estimate consumer willingness to pay accurately from dialogue even when the buyer agent is instructed not to disclose budgets.
Prompt-level confidentiality instructions fail to block the leakage in AI-mediated shopping.
Preference privacy requires interventions at the architectural level rather than through buyer-agent prompts.
Delegation of purchase decisions to language-model agents inherently alters information flows between consumers and sellers.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Consumers may respond by giving AI agents less detailed verbal profiles, which could reduce the personalization benefits the agents are meant to provide.
Market platforms might need new protocols that limit how much identity information is passed through agent communications.
The same leakage pattern could appear in other domains where AI agents represent users through natural language, such as negotiation or service requests.

Load-bearing premise

That the difference in seller inference between the natural-language profile condition and the numeric-budget condition with confidentiality instructions is caused only by role coherence and not by other differences in agent behavior or task interpretation.

What would settle it

An experiment in which buyer agents in the verbal-profile condition receive stronger instructions to mask budget details and seller inference accuracy then falls to the level seen in the numeric-budget condition.

Figures

Figures reproduced from arXiv: 2604.26220 by Salar Nozari, Soogand Alavi.

**Figure 1.** Figure 1: Experimental pipeline. Each buyer–seller interaction produces one transcript; the inference agent reads it to view at source ↗

**Figure 2.** Figure 2: Representative verbal-profile transcript from the vp view at source ↗

**Figure 3.** Figure 3: Inferred versus target willingness to pay, by condition. Each point is a cell mean with bootstrap 95% view at source ↗

read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper flags a plausible privacy leak in AI shopping agents but the experiment is too thinly described to judge whether role coherence is isolated.

read the letter

The main thing to know is that buyer agents given natural-language consumer profiles apparently let sellers recover willingness to pay almost one-for-one from the conversation, and the authors argue this happens through role coherence rather than generic instruction failure. They contrast it with a numeric-budget condition that includes confidentiality instructions and claim the difference shows the effect is tied to the verbal role itself. That framing is useful because it suggests prompt tweaks alone won't solve it and points toward architectural changes instead. The paper does a reasonable job situating the issue in delegation rather than in LLM quirks that might be patched at the instruction level. The comparison setup is a logical first cut for separating the two mechanisms. The soft spots are in the evidence. The abstract supplies no sample sizes, no description of the seller's inference procedure, no prompt templates, and no statistical details, so it is impossible to check whether the two conditions were matched closely enough to rule out prompt-structure artifacts. The stress-test concern lands: if the verbal and numeric prompts differ in length, structure, or context given to the seller model, the gap in recovered WTP could be an artifact rather than proof of role coherence. No code, data, or full methods are referenced here, which leaves the central empirical claim unevaluable. This is for people working on multi-agent systems, AI-mediated commerce, or privacy in language-model agents. A reader already thinking about information flows in automated markets would find the concept worth considering, but only as a prompt for further work rather than a settled result. It deserves peer review because the underlying channel is worth checking and the authors have a clear route to tighten the design, but any referee would need the full experimental protocol and prompt examples before the claim can be assessed.

Referee Report

2 major / 2 minor

Summary. The manuscript argues that delegating purchase decisions to language-model buyer agents supplied with natural-language consumer profiles creates an information channel termed 'role coherence,' enabling seller agents to infer willingness-to-pay (WTP) from dialogue alone. In an experiment contrasting a verbal-profile condition against a numeric-budget condition that includes explicit confidentiality instructions, the authors claim seller-side inference recovers WTP nearly one-for-one in the verbal case while the numeric case does not, thereby isolating role coherence from generic instruction-following failure. The paper concludes that this leakage is inherent to delegation and cannot be mitigated by prompt engineering, proposing instead architectural interventions that trade personalization against privacy.

Significance. If the experimental isolation holds, the result identifies a structural privacy risk in AI-mediated markets that arises directly from the use of natural-language roles rather than from prompt leakage. The controlled LM-to-LM setup provides a reproducible testbed for measuring preference inference, and the distinction between role coherence and instruction failure is a useful conceptual contribution. However, the strength of the claim depends on whether the two buyer conditions are prompt-matched and whether the seller inference procedure is applied identically; absent those controls the result risks being an artifact of experimental design rather than a general property of delegation.

major comments (2)

[Experimental Setup] The central isolation claim (verbal-profile leakage vs. numeric-budget non-leakage) is load-bearing for the paper's argument that role coherence is distinct from instruction-following failure. The abstract states the numeric condition uses 'confidentiality instructions,' but without explicit confirmation in the experimental section that the two buyer prompts are otherwise identical in structure, length, and formatting, any difference in recovered WTP could be driven by prompt template differences rather than the presence/absence of a natural-language role.
[Results] The claim that seller inference 'recovers willingness to pay nearly one-for-one' requires quantitative support (e.g., regression slope, R^{2}, or correlation coefficient) together with sample size, number of trials, and controls for seller-model variation. If these details appear only in an appendix or are omitted from the main results table, the strength of the one-for-one recovery cannot be evaluated.

minor comments (2)

[Abstract] The abstract introduces 'role coherence' without a concise definition; a one-sentence gloss in the abstract would help readers before the full elaboration.
[Notation] Notation for willingness-to-pay (WTP) and recovered estimates should be introduced consistently (e.g., WTP vs. inferred WTP) to avoid ambiguity when the same symbols appear in both conditions.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive feedback. The comments highlight important aspects of experimental controls and result presentation that we will address to strengthen the manuscript. We respond to each major comment below.

read point-by-point responses

Referee: [Experimental Setup] The central isolation claim (verbal-profile leakage vs. numeric-budget non-leakage) is load-bearing for the paper's argument that role coherence is distinct from instruction-following failure. The abstract states the numeric condition uses 'confidentiality instructions,' but without explicit confirmation in the experimental section that the two buyer prompts are otherwise identical in structure, length, and formatting, any difference in recovered WTP could be driven by prompt template differences rather than the presence/absence of a natural-language role.

Authors: We agree that explicit confirmation of prompt equivalence is essential to support the isolation of role coherence from generic instruction-following effects. The experimental section was written with the intent that the two buyer conditions differ solely in profile format (natural-language role versus numeric budget plus confidentiality clause), with all other elements—including overall structure, length, and additional instructions—held constant. To eliminate any ambiguity, we will revise the manuscript to include the complete prompt templates for both conditions in a new table within the experimental setup section and add an explicit statement confirming that the prompts were matched on all dimensions except the role description itself. revision: yes
Referee: [Results] The claim that seller inference 'recovers willingness to pay nearly one-for-one' requires quantitative support (e.g., regression slope, R^{2}, or correlation coefficient) together with sample size, number of trials, and controls for seller-model variation. If these details appear only in an appendix or are omitted from the main results table, the strength of the one-for-one recovery cannot be evaluated.

Authors: We acknowledge that the quantitative backing for the near one-for-one recovery should be more immediately visible in the main text. The results section and associated figures present the regression-based evidence for WTP recovery in the verbal-profile condition (contrasted with the numeric condition), along with details on the number of dialogues and seller-model controls. These supporting statistics are currently summarized in the main results and expanded in the appendix. We will revise by adding a concise table to the main results section that reports the regression slope, R², sample size, number of trials, and seller-model controls, ensuring the strength of the claim can be assessed without reference to the appendix. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical claim rests on experimental comparison, not derivation or self-reference

full rationale

The paper advances an empirical argument from an experiment in which a language-model buyer agent shops on behalf of a verbal consumer profile, with seller-side inference from dialogue recovering willingness-to-pay nearly one-for-one. The abstract and provided text contain no equations, fitted parameters, predictions derived from inputs, or self-citations that bear the central load. The comparison to a numeric-budget confidentiality condition is presented as isolating role coherence, but this is an experimental design choice whose validity is external to any internal reduction; the claim does not reduce to its own inputs by construction. No self-definitional, fitted-input, or uniqueness-imported patterns appear. The result is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that the experimental conditions isolate role coherence and that LLM behavior in the tested setup generalizes to real delegation.

axioms (1)

domain assumption LLM buyer agents will follow role instructions in a way that preserves coherence with the provided consumer profile.
Invoked to explain why natural-language profiles create leakage while numeric budgets do not.

invented entities (1)

role coherence no independent evidence
purpose: Information channel through which sellers infer WTP from buyer-agent dialogue.
New term introduced to name the leakage mechanism arising from delegation itself.

pith-pipeline@v0.9.0 · 5425 in / 1241 out tokens · 36843 ms · 2026-05-07T12:56:39.157396+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Acquisti A, Varian HR (2005) Conditioning prices on purchase history.Marketing Science24(3):367–381. Allouah A, Besbes O, Figueroa JD, Kanoria Y , Kumar A (2026) What is your ai agent buying? evaluation, biases, model dependence, & emerging implications of agentic e-commerce.Proceedings of the ACM Web Conference 2026, 8697–8700. Andreas J (2022) Language ...

work page arXiv 2005
[2]

Aditya Singh, Gerson Kroiz, Senthooran Rajamanoharan, and Neel Nanda

Cherep M, Ma C, Xu A, Shaked M, Maes P, Singh N (2025) A framework for studying ai agent behavior: Evidence from consumer choice experiments.arXiv preprint arXiv:2509.25609. Dub´e JP, Misra S (2023) Personalized pricing and consumer welfare.Journal of Political Economy131(1):131–189. Gopal RD, Li J, Riemer K, Sarker S, Singh PV , Susarla A, Bichler M, Tha...

work page doi:10.1038/s41586-023-06647-8 2025