The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets

AI agents are increasingly used in consumer applications for product search, negotiation, and transactions. We investigate a setting where both consumers and merchants authorize AI agents to automate negotiations and transactions. We address two questions: (1) Do different LLM agents exhibit varying performance when making deals for users? (2) What are the risks when using AI agents to fully automate negotiations in consumer settings? We design an experimental framework to evaluate AI agents’ capabilities in real-world negotiation scenarios, experimenting with various open-source and closed-source LLMs. Our analysis reveals that deal-making with LLM agents is an inherently imbalanced game. Furthermore, LLMs’ behavioral anomalies might lead to financial losses for both consumers and merchants through overspending or unreasonable deals. While automation can enhance efficiency, it poses significant risks to consumer markets. Users should be cautious when delegating business decisions to LLM agents. All the code and data are available at https://github.com/ShenzheZhu/A2A-NT.

Introduction. Business negotiation and deal-making lie at the heart of the modern economy, yet achieving agreement is rarely straightforward. It requires effective information gathering, strategic reasoning, and skilled negotiation and decision-making (Lewicki and Hiam, 2011; Agndal et al., 2017). Recently, large language model (LLM) powered AI agents have demonstrated remarkable capabilities and are increasingly adopted for real-world tasks (Xu et al., 2024; Masterman et al., 2024). Given the importance of negotiation in business operations, researchers and practitioners have begun exploring ways to leverage AI agents to automate shopping and sales processes for both consumers and merchants (Kong et al., 2025; Chen et al., 2024), mostly assuming agents interact with real human users. However, with rapid AI agent adoption in consumer markets, both consumers and merchants might delegate their negotiation and decision-making to AI agents and direct agent-to-agent interactions might soon be commonplace. Given the natural capability differences of AI agents in negotiation settings (Bianchi et al., 2024) and unique agent-to-agent negotiation dynamics (Vaccaro et al., 2025), it becomes a key question: What happens when consumers and merchants use AI Agents with different capabilities to automate their negotiation and transactions in consumer settings? In this study, we propose a comprehensive framework to investigate opportunities and risks associated with fully automated, user-authorized agent-toagent negotiation and transaction. Inspired by realworld shopping and sales workflows, we design an experimental setting where a buyer agent attempts to negotiate lower prices based on user-defined budgets, while a seller agent, aware of wholesale costs, aims to maximize profit. Each agent independently makes decisions throughout negotiation, simulating fully autonomous, end-to-end transactions between AI agents. To evaluate negotiation behaviors and capabilities of AI agents in realistic consumer scenarios, we compile a dataset of 100 real-world products across three major categories: electronic devices, motor vehicles, and real estate. These products vary in nature and price range, reflecting diverse consumer transactions. For each item, we collected actual retail prices and estimated wholesale values, which were provided to seller agents to simulate authentic market dynamics. We conducted negotiation and transaction experiments using several advanced language models, including GPT series (Hurst et al., 2024), Qwen-2.5 series (Yang et al., 2024a), and DeepSeek series (Liu et al., 2024; Guo et al., 2025). Our analysis reveals substantial negotiation performance gaps across models that correlate with their general capabilities and specifications. More capable models consistently secure better deals as both buyers and sellers. This suggests that in real-world scenarios, parties using less capable AI agents would face systematic economic disadvantages and financial losses. Beyond performance differences, we identify several key risks associated with delegating negotiation and transactional authority to AI agents: (1) Constraint violation risk: Buyer agents may disregard user-imposed budget constraints, completing purchases users cannot afford. Similarly, seller agents may accept prices below wholesale costs, leading to financial losses; (2) Excessive payment risk: buyer agents sometimes offer higher prices than retail price, resulting in unnecessary overpayment; (3) Negotiation deadlock risk: agents may become stuck in prolonged negotiation loops without reaching agreement; (4) Early settlement risk: higher budget settings lead buyer agents to compromise more readily, instead of striving for better deals. This contrasts with low-budget scenarios, where agents demonstrate stronger price resistance and negotiation effort. These findings have important implications for agent-assisted decision-making in consumer markets. Access to more powerful AI models can lead to better deals, potentially reinforcing economic disparities among users. Furthermore, weaknesses in LLMs, such as limited numerical reasoning and occasional failures in instruction-following, can expose both consumers and businesses to systemic financial risks. As fully autonomous agentto-agent interactions become more common, practitioners should exercise caution when delegating high-stakes decisions to AI agents. This paper makes the following contributions:

Method. 2 Modeling Agent-to-Agent Negotiations and Transactions The goal of this paper is to systematically investigate outcomes and risks when AI agents are authorized to negotiate and make decisions on behalf of consumers and business owners. To this end, we introduce an experimental setting that closely reflects real-world negotiation and transaction scenarios in consumer markets. More specifically, we instruct LLM agents to engage in price negotiations over real consumer products, with one agent acting as buyer and the other as seller. By observing model behaviors in these structured and realistic scenarios, we aim to forecast potential behaviors, strategies, and risks that may arise as such agentmediated transactions become more prevalent in future consumer environments.

2.1 Basic Notations and Definition We define the key symbols used in this paper. The total number of negotiation rounds is denoted as T, which may be fixed or dynamically inferred. Let pr be the retail price, pw be the wholesale price, β be the buyer’s budget, and φ be the product features. The proposed price pa at round t is pt a, and the price trajectory is P = {pt a}T t=1 with pT a as the final round proposed price1.

2.2 Negotiation Scenario 2.3 Negotiation Pipeline from trustworthy sources. As the wholesale cost pw may not be directly available on the public internet, we prompt GPT-4o with item-specific information and current market conditions to estimate a reasonable wholesale cost pw based on industry norms. More details of dataset creation are shown in Appendix C.

Discussion. 4 Anomaly-Induced High-Stakes Risks Autonomous AI agents could potentially bring huge economic value to the users in many settings. However, they may also introduce systematic risks when being deployed at large (Feliu, 2001; Jabłonowska et al., 2018; Rohden and Zeferino, 2023; Deng et al., 2025; Hammond et al., 2025; Chen et al., 2025). In this section, we discuss the potential risks when both buyers and sellers delegate their negotiations and decision-making to AI agents and how models’ anomalies may translate into tangible economic losses for real users. We 4.1 From Model Anomaly to Financial Risks Fully automated, agent-based negotiation systems are prone to various anomalies stemming from unstable decision-making and imperfect instruction following of their base LLMs (Lan et al., 2025; Zhang et al., 2025; Cemri et al., 2025). While such failures may seem trivial or expected in research settings, they pose tangible risks to users in real-world settings. In this section, we analyze four model behavioral anomalies, pinpoint the conditions that trigger them, and outline how they can be translated into real financial loss for users. The detailed mathematical formula for the following anomaly measurement can be found in Appendix D.2.

Constraint Violation. Consider scenarios where a user authorizes an AI agent to negotiate with a fixed budget β. If the agent accepts a deal above the budget, it may overdraw the account or exceed the user’s willingness to pay. Similarly, a seller agent agreeing to prices below cost pw incurs guaranteed losses. We quantify such anomalies using Out-of-Budget Rate (OBR) and Out-of-Wholesale Rate (OWR). As shown in Figure 6, models with stronger negotiation capabilities, such as DeepSeek series and Latest Generation GPT Series (GPT-4.1, deals in low-budget scenarios. This suggests that while instruction-following failures are often considered trivial, they can pose serious financial risks to both buyers and sellers in real consumer settings.

Excessive payment. Our experiments uncover a surprising anomaly: buyer agents sometimes pay more than the listed retail price. We quantify this behavior with Overpayment Rate (OPR), the proportion of successful deals where the final transaction price exceeds the retail price despite the buyer’s budget allowing a lower amount. As shown in Figure 8 (top), overpayment frequently occurs under high-budget settings. Except for the DeepSeek family and Latest Generation GPT Series (GPT-4.1, o4-mini and o3), every model overpays when buyers have large β values. We qualitatively examine negotiation histories and found that overpayment often occurs after sellers ask buyers to reveal their budget early in conversations. Despite our system prompt explicitly instructing buyers not to disclose their budget unless necessary, many buyer agents reveal their budget easily. Sellers then anchor offers to the disclosed number, even when higher than the listing price, and buyers accept the inflated deal without objection.

Negotiation Deadlock. Imagine a user who uses an API-based buyer agent for negotiation, expecting it to operate efficiently within reasonable bounds. Since users are billed per token or API call, they assume the agent will either reach a deal or end negotiation appropriately. However, we observe that agents might continue bargaining even when sellers state firm bottom lines, leading to unnecessarily long negotiations. This wastes computational resources and undermines automation’s practical utility. We define this issue as "Negotiation Deadlock," formally defined as any dialogue reaching maximum rounds Tmax without final agreement or explicit rejection.

Conclusion. With large-scale deployment of AI agents in real consumer settings, agent-to-agent interactions will become ubiquitous in the near future. But what happens when we fully automate negotiation and deal-making with consumer and seller authorized AI agents? In this paper, we design an experimental framework to investigate potential issues and risks in agent-to-agent negotiations and transactions. Our analysis reveals that agent-to-agent negotiation is naturally an imbalanced game where users with less capable agents face significant financial loss against stronger agents. Furthermore, we found that LLMs’ anomalies might transfer to real economic loss when deployed in real consumer settings. Our paper highlights the potential risks

Limitations. 6 Limitation