LLM-Agent Interactions on Markets with Information Asymmetries

As AI agents increasingly act on behalf of human stakeholders in economic settings, understanding their behavior in complex market environments becomes critical. This article examines how Large Language Models coordinate on markets that are characterized by information asymmetries and in which providers of services have incentives to exploit that asymmetry for their own economic gain. To that end, we conduct simulations with GPT-5.1 agents in credence goods markets, manipulating the institutional framework (free market, verifiability, liability), LLM agent’s social preferences (default, self-interested, inequity-averse, efficiency-loving), and reputation mechanisms across one-shot and repeated 16-round interactions. In one-shot settings, LLM agents largely fail to establish cooperation, with markets breaking down except under liability rules or when experts have efficiency-loving preferences. Repeated interactions solve consumer participation through competitive price reduction, but expert fraud remains entrenched absent explicit other-regarding preferences. LLM consumers focus narrowly on price levels rather than understanding strategic incentives embedded in markups, making them vulnerable to exploitation. Compared to human experiments, LLM markets exhibit substantially higher consumer participation but much greater market concentration, lower prices, and more polarized fraud patterns. The effect of institutions like verifiability and reputation is also much more ambiguous. Surplus shifts dramatically toward consumers under social-preference objectives. These findings suggest that institutional design for AI agent markets requires fundamentally different approaches than those effective for human actors, with social preference alignment emerging as the primary determinant of market efficiency.

Introduction. Recent advances in (generative) artificial intelligence (AI) and strategic foundation models suggest a world in which humans will increasingly rely on and delegate to autonomous AI agents (Goktas et al. 2025; Imas, Lee, and Misra 2025; Ivanov et al. 2024). Already today it is easy to imagine AI agents representing and acting on behalf of human stakeholders on, for example, online marketplaces, shopping platforms, or when negotiating (digital) services.

In this paper, we examine current LLM (Large Language Model) capabilities to navigate relatively complex market structures that are characterized by information asymmetries between agents: credence goods markets (or markets for expert services). Credence goods markets are important, because many important economic and social domains are characterized by information advantages where superiorly informed sellers or suppliers (here: experts) can exploit their informational advantage at the expense of consumers (Dulleck and Kerschbamer 2006). Prominent examples include financial services, expert advice, medical services, or legal services. Over the last 20 years, a lot of empirical field research has revealed strong real-world inefficiencies, including widespread rates of overcharging (i.e. charging for more services than provided) and under-treatment (i.e. not solving the consumer’s problem or not satisfying the consumer’s demand)(Gottschalk, Mimra, and Waibel 2020; Balafoutas et al. 2013; Kerschbamer, Neururer, and Sutter 2016, 2023; Balafoutas and Kerschbamer 2020; Schneider 2012). Importantly, these problems are difficult to solve, because the information asymmetry precludes consumers from disciplining experts through informed selection, and effective regulations such as liability are often difficult or impossible to enforce (it is, for example, often impossible to prove expert fraud or negligence). AI agents, however, may have the potential to alleviate these information problems.

In contrast to human actors, current foundation models like LLMs have access to vast amounts of expert knowledge. They are able to think strategically at scale, and learn from real-time feedback. Recent work argues that human cognitive and informational restraints may be a key driver of AI agent adoption (Shahidi et al. 2025). On the expert’s side, there are already numerous examples of automated LLM-provided services, ranging from financial services to medical advice, taxation services, and shopping assistants (Ayers et al. 2023; Hirosawa et al. 2023; Nov, Singh, and Mann 2023; Zhao et al. 2024; Li et al. 2023; Oehler and Horn 2024; Lai et al. 2024; Nay et al. 2024; Alarie, Niblett, and Yoon 2018; Reis, Reis, and Kunde 2024; Shekar et al. 2024; Ayre, Cvejic, and McCaffery 2024; Leslie-Miller et al. 2024; Seabrooke et al. 2024; Schneiders et al. 2024). LLM providers such as OpenAI and Anthropic offer explicit agent services, scaling up consumer access to LLM agents.

However, how well AI agents based on Large Language Models coordinate on markets that operate under information asymmetry is an open question. There is considerable debate about the capacity of AI agents such as LLMs – especially off-the-shelve ones – to navigate complex economic environments (Felin and Holweg 2024; Valmeekam et al. 2023; Raman et al. 2024), with several recent successful examples (Fish, Gonczarowski, and Shorrer 2024; Deshpande and Jacobson 2026; Lopez-Lira 2025; Imas, Lee, and Misra 2025). Laboratory experiments have consistently shown that humans struggle with the game theory of credence goods markets, but may still achieve moderate levels of cooperation through social preferences. AI agents could out-perform human stakeholders through strategic price-setting (experts) and expectation forming (consumers) that alleviate the threat of market unraveling. On the other hand, they may lack the advantage of mutually beneficial (pro-) social preferences that allow humans to enter a reciprocal relationship via initial trust (see e.g., Erlei 2025; Dvorak et al. 2025; Erlei and Rau 2026). In this article, we present a sequence of simulations using GPT-5.1 in which LLM experts interact with LLM consumers on standard experimental credence goods markets (Dulleck, Kerschbamer, and Sutter 2011). We consider both one-shot and repeatedinteraction market scenarios, and vary the institutional market framework (No Institution vs. Verifiability vs. Liability), the LLMs social preferences (default model, self-interested, inequity-averse, efficiency-loving), as well as the possibility of bilateral expert–consumer reputation building to answer the following question:

• How do LLM AI agents coordinate on credence goods markets?

• How do social preferences and reputation affect agent behavior and market welfare?

Related work. This article is situated at the intersection of the literature on credence goods and the literature about LLM agents in strategic economic settings. For human interactions, there exists a large literature of laboratory and online experiments that gauge the efficiency of credence goods markets across various institutional settings (Balafoutas and Kerschbamer 2020; Dulleck, Kerschbamer, and Sutter 2011; Balafoutas et al. 2023; Kerschbamer, Sutter, and Dulleck 2017; Grosskopf and Sarin 2010; Beck et al. 2013; Balafoutas et al. 2015; Kerschbamer, Sutter, and Dulleck 2017; Inderst, Khalmetski, and Ockenfels 2019; Kandul, Lanz, and Reins 2023; Tracy, Schniter, and Zíka 2023; Mimra, Rasch, and Waibel 2016; Agarwal, Liu, and Prasad 2019; Erlei and Meub 2024; Kerschbamer, Neururer, and Sutter 2023; Schneider, Meub, and Bizer 2021). Recently, Erlei (2025) has shown how pro-social human experts can out-compete self-interested human experts by delegating their choices to an LLM agent with a transparent, codified objective function. Furthermore, they show that in the absence of transparent objective functions, human consumers trust LLM experts less than human experts.

Since the introduction of LLMs, there has been a strong increase of articles studying the strategic capabilities and tendencies of AI agents in economic settings (Zhang et al.

2024; Kasberger et al. 2023; Guo 2023; Schmidt et al. 2024; Rahwan et al. 2019). Deshpande and Jacobson (2026) study large language models in oligopolistic Cournot markets. LLM agents are capable of strategic decision-making, and exhibit tendencies for tacit collusion (which has also been observed by Fish, Gonczarowski, and Shorrer (2024)). Imas, Lee, and Misra (2025) allow humans to use LLM agents as instructed agents in a negotiation setting, showing how (1) human personality traits affect agent behavior through instructions, (2) social norms are replicated in agent interactions via, e.g., the prevalence of “fair splits”, and (3) variance in outcomes increases. In Cohen et al. (2025), endowing LLMs with personality traits significantly changes the behavioral dynamics in a negotiation settings, and in Long et al. (2025), emotional expressions by AI agents affect economic outcomes. These patterns are similar to our study, where we endow LLM agents with different objective functions mirroring social preferences. Shah et al. (2025) replicate several classic auction findings from human experiments using LLM agents. They also find

Method. 3. The Market We analyze a standard credence goods problem as a (1) one-shot and (2) repeated-round market (Dulleck, Kerschbamer, and Sutter 2011). Four experts compete over four consumers. Consumers have a big problem with probability h = 0.5 or a small problem with probability 1 −h = 0.5. To solve their problem, consumers need to approach an expert.

After approaching an expert, they are committed to receive the recommended treatment under the offered price. If the problem is solved, consumers receive a payoff V = 10. If the problem is not solved, they earn nothing, and still pay their expert, resulting in a net negative income. Consumers may choose to leave the market untreated and earn an outside option σ = 1.6. Experts who attract no consumer earn 0. By assumption, indifferent consumers visit an expert and experts who are indifferent between dishonest and honest behavior choose to be honest. Experts receive a costless diagnostic signal with 100% accuracy for each consumer. Then, depending on treatment, they first choose between a high cost treatment (HCT) and a low cost treatment (LCT), and subsequently charge consumers for their service. The HCT solves both problems and costs the expert ̄c = 6, while the LCT only solves the small problem and costs the expert ̄c = 2. Experts set a price pair P = ( ̄p, ̄ p) at the beginning of each round where p ∈{1,...,11}, with ̄p (HCT) ≥ ̄ p.

Furthermore, expert decision-making may be constrained by the institutional setting. We consider three cases:

No Institution. There are no constraints on expert behavior. Experts are allowed to freely choose between the two treatments (HCT, LCT) and the two prices ( ̄p, ̄ p) after diagnosing consumers. Consumers know about the expert’s choice set. There are three ways to defraud consumers: under-treatment (LCT if problem is big), over-treatment (HCT if problem is small), over-charging ( ̄p after choosing the LCT).

Verifiability. The implemented treatment is verifiable to consumers. Therefore, experts must charge the price of their chosen treatment. If an expert chooses the HCT, they must choose ̄p, if they choose the LCT, they must choose ̄ p. There are two kinds of fraud: under-treatment, over-treatment.

Liability. Experts are liable and must solve the consumer’s problem. If an expert diagnoses a big problem, they must choose the HCT. If an expert diagnoses a small problem, they can freely choose between the HCT and the LCT. Furthermore, experts are free to choose either price, irrespective of their treatment choice. There are two ways to defraud consumers: over-treatment, over-charging.

In the one-shot simulations, there is only one interaction, and we do not manipulate reputation because there is no future. In the repeated paradigm, experts and consumers play for 16 rounds. Here, we differentiate between reputation and no reputation. In no reputation, consumers cannot identify individual experts across rounds, and thus, in case of mis-treatment, cannot identify the dishonest expert for the subsequent choice. In reputation, experts are identifiable, and can therefore build a reputation for being (dis-)honest with each individual consumer who approaches them. Reputation is always a private signal, not a public signal.

3.1. Predictions We first consider self-interested risk-neutral experts (Dulleck and Kerschbamer 2006) in the one-shot setting (Erlei 2025). Theoretically, little changes under repeated interactions.

Experts compete to attract consumers. In No Institution, the dominant payoff-maximizing expert strategy for each approaching consumer is to choose the LCT and charge the big price ̄p. Consumers anticipate this, and condition their approach choice solely on ̄p, because they expect over-charging. Expected payoff is πc ni = (1 −h) ∗(V − ̄p) −h ̄p. Given their outside option σ = 1.6, the consumer approaches an expert if ̄p ≤3. Because experts have an outside option of σ = 0, always choose the LCT, and always charge ̄p, experts compete over consumers via prices as long as πe ni = ̄p − ̄c > 0, and therefore, ̄p ≥3. Hence, all experts choose P = { ̄ p,3}1, which offers consumers πc ni = 2 > 1.6.

Discussion. Implications. Our results highlight how information asymmetries that induce incentives for fraud remain widespread in agent-based interactions. Even more so, they may become entrenched parts of repeated interactions. Markets converge towards states with rampant fraud that is compensated by somewhat low prices. In reality, this may be difficult to accept for human stakeholders. However, it does lead to sustained expert-consumer interactions, which – even when self-interested experts consistently intend to defraud consumers – strongly outperforms one-shot settings and reaches efficiency rates of roughly 31% on average. Notably, market efficiency climbs to 88% with efficiency-loving objective prompts and 83% for inequity-averse prompts. This coincides with strong expert losses, leading to negative aggregate income in efficiency-loving × No Institution, except for inequity-averse experts under verifiable treatment outcomes. Therefore, Verifiability provides an institutional structure that rewards experts for considering consumer outcomes. This may be one promising avenue when designing agentic credence goods markets. Similarly, private reputation institutions like verified identities appear more consistently useful under verifiability, whereas in No Institution, the effects are very ambiguous, often harming overall welfare. In conclusion, under complicated market structures that involve information asymmetries and incentives for fraud, LLM behavioral structurally diverges from standard theory, leading to novel empirical patterns like entrenched fraud. Social preferences can alleviate dishonesty, but often at the cost of expert welfare. Verifiability provides some promise, as it is the only environment that offers experts incentives to deviate from self-interest, and may positively affect the value of private reputation institutions that reduce under-treatment and over-charging. However, in many conditions, verifiability consistently worsens market outcomes, suggesting that it is far from a panacea. In gen-

Conclusion. This article provides initial results about economic outcomes and behavioral trends in agentic credence goods markets. We find that LLM agents do not follow the strategic behavior dictated by the standard model. For one-shot settings, this causes market breakdown in the absence of liability. In the repeated setting, LLM experts get away with widespread consumer fraud because LLM consumers appear to condition their behavior mostly on prices. In comparison to human data, market concentration increases, and the welfare effects of institutions becomes much more ambiguous.

One-Shot. LLM Agents (here GPT-5.1) do not consider incentives through markups in their pricing choices and therefore regularly offer negative expected value deals to consumers. This leads to widespread market breakdown. However, even when prices follow theory (e.g., equal markups in self-interested × verifiability), LLM consumers do not approach LLM experts because they do not trust indifferent experts to not defraud them. AI-agents behave dishonestly by default and when prompted to follow the objective of maximizing their own economic position. Inducing other-regarding preferences alleviates LLM dishonesty, but does not improve price-setting. Notably, efficiency-loving LLM experts consistently set prices that are so low that they always attract consumers but hurt the expert’s income through negative returns. Apparently, the LLM prioritizes consumer participation, which is indeed the most important factor for aggregate welfare. However, because experts lose on average, this is not a stable strategy without some re-distributive institution. Compared to recent behavior evidence from one-shot experiments with humans (Erlei 2025), LLM agents are generally substantially less successful than humans in establishing cooperation on credence goods markets.