Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation

As AI usage becomes more prevalent in social contexts, understanding agent-user interaction is critical to designing systems that improve both individual and group outcomes. We present an online behavioral experiment (N= 243) in which participants play three multi-turn bargaining games in groups of three. Each game, presented in randomized order, grants access to a single LLM assistance modality: proactive recommendations from an Advisor, reactive feedback from a Coach, or autonomous execution by a Delegate. All three modalities are powered by an LLM with super-human performance within this negotiation setting. On each turn, participants privately decide whether to act manually or use the AI modality available in that game. We document a preference–performance misalignment: participants strongly prefer the higher-control Advisor (44%) over the Delegate (19%), yet groups only significantly increase collective surplus under Delegate access. Adjusting for voluntary non-compliance, delegating to the AI yields suggestive individual welfare gains, roughly 1.5× the intent-to-treat estimate. A mechanism analysis traces this gap to a human filter: AI-generated proposals create more joint surplus than manual proposals across all conditions, but in the Advisor and Coach modes users modify, override, or ignore the AI’s suggestions, reverting toward human-baseline trade patterns. The Delegate advantage arises not from a different AI capability but from bypassing this filtering step altogether. Realizing these welfare gains depends not only on model capability, but on the interaction structure through which that capability is delivered. We argue that assistance modalities should be designed as mechanisms with endogenous participation; adoption-compatible interaction rules are a prerequisite to improving welfare with automated assistance.

Introduction. Large language models (LLMs) are shifting from passive tools to autonomous agents, capable of navigating complex social tasks and reshaping the incentives of multi-party environments. Understanding the design and effects of such systems is key to the emerging “agentic economy” [Rothschild et al. 2025; Shahidi et al. 2025; Tomasev et al. 2025]; furthermore, in shared environments, how much agency each stakeholder cedes to an AI directly shapes collective welfare alongside individual outcomes. The tension in delegating actions to autonomous agents is increasingly operationalized through a diverse spectrum of assistance modalities. Beyond traditional chat interfaces that passively respond to user input, emerging systems range from proactive assistants that initiate guidance [Google 2025; OpenAI 2025b] to autonomous agents empowered with independent execution [OpenAI 2025a]. However, the impact of these design choices on strategic outcomes—and their scalability in multi-party environments—remains underexplored. Prior research on human-AI collaboration has primarily evaluated single-user, single-AI interactions in domains such as medical diagnosis, credit assessment, and risk prediction [Agarwal et al. 2023; Bansal et al. 2019; Green and Chen 2019a], often targeting objectives with verifiable ground truths [Vaccaro et al. 2024]. These studies have established foundational insights, such as a preference-performance misalignment, where users may prefer sub-optimal agents that require less cognitive load [Bansal et al. 2019; Buçinca et al. 2020, 2021]. However, virtually all empirical evidence on human–AI interaction comes from individual decision-making tasks—settings where one user interacts with one AI system and the outcome depends only on that dyad. Experimental data on AI assistance in multi-party, strategically interdependent environments—where one participant’s AI-assisted action changes the payoff landscape for others—remains scarce. This gap matters because multi-party settings introduce externalities, equilibrium effects, and adoption spillovers that cannot be studied in isolated dyads. Understanding these dynamics requires moving beyond individual users to examine how AI-mediated technology reshapes collaborative group outcomes [Grudin 1988]. Such human-AI interaction patterns become increasingly complex in social, real-world applications, which are defined by strategic interdependence, dynamic equilibria, and collective externalities. Recent work on autonomous LLM negotiations demonstrate that agent-to-agent bargaining can exhibit distinct risks and behaviors relative to human negotiation [Qian et al. 2025a; Zhu et al. 2025]. A complementary line of research studies principal-agent interaction patterns in social contexts, where humans customize an agent (often via prompt writing) to negotiate autonomously on their behalf [Imas et al. 2025; Vaccaro et al. 2025]. While these studies establish the feasibility of autonomous negotiation, they abstract away from the practical design choice of user controllability—specifically, how different modalities of interaction influence adoption and equilibrium outcomes.1 We study how different allocations of agency affect behavior and welfare in a strategically interdependent setting, operationalizing the spectrum of AI agency into three distinct interaction modalities while holding the underlying model capability constant: an Advisor (proactive recommendations), a Coach (reactive feedback), and a Delegate (autonomous action). These three modalities are grounded in the levels-of-automation framework proposed by Sheridan and Verplank [1978] and formalized by Parasuraman et al. [2000], who characterize a continuum from full human control to full machine autonomy. Rather than sampling arbitrarily along this continuum, our design selects three qualitatively distinct allocations of decision authority: the Advisor initiates action and the human retains veto power; the Coach inverts this flow, with the human initiating and the system critiquing before execution; and the Delegate acts autonomously on the human’s behalf. These three points represent the principal ways in which initiative and oversight can be partitioned between a human and an AI assistant: AI-proposes/human-decides, human-proposes/AI-critiques, and AI-decides/human-observes.2 We evaluate these modalities through a bargaining game [Qian et al. 2025a] designed to evaluate AI capabilities in group negotiations, a setting with empirically exhibited human inefficiencies in trading behaviors [Binmore et al. 1986; Rubinstein 1982]. Our LLM agents utilize prompt scaffolding based on Gemini-2.5 Flash [Comanici et al. 2025] that empirically outperforms human traders in the All-Agent Baseline, allowing us to isolate the effects of the interaction structure.

Related work. LLM-based agents demonstrate increasingly strong social capabilities across negotiation, persuasion, mediation, and multi-agent coordination tasks [Abdelnabi et al. 2023; Agashe et al. 2023; Horton Choose Your Agent: Tradeoffs in Adopting AI Advisors, Coaches, and Delegates in Multi-Party Negotiation 0:5 LLM negotiators through prompt-designed competitions [Imas et al. 2025; Vaccaro et al. 2025] and agent–agent market simulations [Zhu et al. 2025]. While these studies primarily evaluate agent capabilities in bargaining environments, our study evaluates the human adoption of such capabilities through varying interaction modalities.

2.2 Delegation in Human–AI Interaction Full delegation to agentic systems remains relatively novel in empirical studies of human–AI teaming; a recent meta-analysis [Vaccaro et al. 2024] reviewed over 100 experiments and found that only a small subset involve structured delegation of decision authority. When delegation is studied, it typically appears in isolated tasks where complementarity can be engineered; for example, hybrid human–AI assessments in physical therapy [Lee et al. 2021], post-editing pipelines in summarization [Lai et al. 2022], or selective triage allocations based on uncertainty or rule-based heuristics [Agarwal et al. 2025b; Raghu et al. 2019]. In group settings, agent delegation can increase cooperation in collective-risk dilemmas [Domingos et al. 2021]. AI-mediated group consensus can be perceived as clearer and less biased [Tessler et al. 2024]. In multi-party negotiation, AI moderation can raise fairness and efficiency perceptions [Kobiella et al. 2025]. However, delegating shared tasks can disrupt the mutual cooperative awareness that groups use to coordinate actions and infer counterpart intent [Dourish and Bellotti 1992]. This introduces a mismatch between computational optimization and the flexible, norm-sensitive realities of human group work [Ackerman 2000]. We compare voluntary delegation against advisory and coaching modalities, evaluating how ceding control to an agent shapes cooperative dynamics and welfare outcomes in an interdependent game.

2.3 Adoption and Control in AI Assistance Despite potential benefits of AI adoption and delegation, takeup of such systems by human users often involves a calculated tradeoff between outcome quality and perceived control [Owens et al. 2014; Ryan and Deci 2000]. Human–AI teams may fail to achieve synergy even when AI improves accuracy [Green and Chen 2019a; Vaccaro et al. 2024], in part because designs that enhance objective performance may increase cognitive effort or monitoring costs [Buçinca et al. 2020, 2021; Davis 1989; Qian and Wexler 2024; Sweller 1988]. Users must also maintain accurate mental models of when AI is likely to err [Bansal et al.

Method. 3 Experiment We frame our empirical inquiry through three sets of research questions. Broadly, when access to AI is given to individuals in a multi-user context, who uses it, what do they gain, and what mechanism drives the welfare effects?

RQ1: Welfare effects. At the group level, does access and adoption of LLM assistance—in any of the three modalities (Advisor, Coach, Delegate)—increase individual- and group-level surplus, relative to a no-AI baseline?

RQ2: Adoption and preferences. How do participants’ stated preferences and usage decisions vary across the Advisor, Coach, and Delegate modalities, independent of their welfare effects?

RQ3: Mechanisms. What mechanisms explain the differential welfare effects across modalities? Specifically, how does human intervention in Advisor and Coach modes alter the quality of AIgenerated proposals, and to what extent does this filtering account for the Delegate advantage?

We explore the negotiation dynamics in the context of a chip bargaining game introduced by Qian et al. [2025b], where participants take turns exchanging chips of different colors, with randomly assigned private valuations, with the goal of maximizing their surplus. During each turn, one participant proposes an offer (e.g. 7 red for 3 green chips), and the other two simultaneously and privately decide whether to accept or decline. If both accept, one is chosen randomly to clear the trade. The game ends after nine turns; participants leave with any surplus earned beyond the initial value of their chips. This stylized resource-exchange game models real-world collaborative tensions— cross-departmental scheduling, budget negotiations, and multi-party project coordination—where players must find Pareto-improving allocations under asymmetric information and conflicting interests. This environment provides several properties that enable our investigation. First, by endowing participants with pre-defined preferences, it provides an objective ground truth for performance and outcome measurement, which is a requirement in empirical trading games exploring similar behaviors [Bochet et al. 2024; Kahneman et al. 1990; Roth et al. 1991]. Second, the combination of information asymmetry and restricted communication create a setting that necessitates strategic reasoning.3 Finally, because players’ outcomes are coupled through the channels detailed below, the multi-player, non-zero-sum design permits direct measurement of the group-level externalities that individual-task designs cannot isolate.

Strategic interdependence. Most empirical evidence on human–AI collaboration, and in particular the bulk of the delegation and algorithm-aversion literature, comes from individual decision-making tasks in which one user interacts with one AI system and the outcome depends only on that dyad [Agarwal et al. 2023; Bansal et al. 2019; Dietvorst et al. 2015; Green and Chen 2019a; Vaccaro et al. 2024]. Canonical domains—medical diagnosis, credit scoring, forecasting, deception detection— admit a single ground-truth target that the user and the assistant jointly estimate, and the assistance modality affects only the decision-maker’s own payoff. Our game departs from this class in three specific ways:

• Payoff coupling. A player can realize surplus only through trades that a counterparty agrees to execute; each player’s final payoff is produced jointly with at least one other seat. • Opportunity-set externality. Every consummated trade permanently alters the inventories of the two trading players, which shrinks or expands the set of feasible trades available on every subsequent turn—including the turns of the non-participating third player. • Informational externality. Every proposal and every accept/decline decision reveals partial information about the actor’s private valuations. All three players update beliefs from the public trade history, so one player’s action shifts the best responses of the other two for the remainder of the game.

Discussion. Aggregate performance metrics can obscure how groups achieve outcomes and who benefits. We analyze our results through three core themes: how voluntary AI delegation represents a social dilemma of adoption; how the “human filter” acts as a socio-technical translator between raw efficiency and social norms; and how task structure and organizational hierarchy shape the optimal allocation of agency.

5.1 Socio-Technical Dynamics and the Inverted Collaboration Paradox In collaborative systems research, Grudin’s Paradox posits that technology fails to adopt because the individuals who bear the cost of system entry (e.g., manual data entry) receive zero individual utility, while others reap the benefits [Grudin 1988]. Our findings with the Delegate modality present a distinct, structural variation of this pattern. Because Delegate adopters receive suggestive, individual-level welfare gains ( ˆβ= 0.043, Table 1), delegation is materially a win-win scenario rather than a classic utility misalignment. The barrier to adoption is not structural (a lack of individual utility), but psychological: the delegating participant cedes agency and bears the psychological cost of relinquishing decision control, yet their adoption generates positive group externalities that benefit counterparts. While the group-level ITT trend is suggestive ( ˆβ= 0.084, p= 0.033, padj = 0.10), our tradelevel decomposition in Section 4.3 reveals a powerful market-wide spillover: manual proposals in Delegate-access games yield higher joint surplus (3.110) than those in Advisor (2.678) or Coach (2.746) games (p< 0.01). The Delegate acts as a local “market maker,” introducing higher-quality, non-conservative proposals into the shared offer pool, lifting the welfare floor for adopters and non-adopters alike. Crucially, this spillover provides an empirical example against the central ethical concern of algorithmic exploitation in multi-party systems: that an optimization agent might maximize joint surplus by extracting value from unassisted counterparts through highly asymmetric splits. In our study, unassisted counterparts (non-adopters) in Delegate games actually achieved a higher average individual surplus (0.218) than the active AI adopters who utilized the Delegate (0.201). This empirical fact provides an instance of transactional equity: rather than acting as a predatory tool that exploits counterparts, the Delegate agent upgrades the shared decision space, permitting counterparts to capture equal or even greater welfare gains. Structurally, this creates a social dilemma of adoption: individual-level algorithm aversion and a desire for decision agency prevent the group from reaching a system-level, Pareto-efficient cooperative equilibrium. Designing collaborative systems in these settings is therefore a mechanism design problem: we must design interfaces and interaction rules that reduce these psychological ceding costs to unlock group-level welfare outcomes [Grudin 1988]. Why does delegation outperform advisory and coaching interfaces when using the identical underlying model? In the Advisor and Coach conditions, human intervention acts as a human filter: users bypass the AI, edit proposals toward conservative allocations, or ignore coaching feedback. In mixed-motive bargaining, the conservative, symmetric proposals that humans favor (such as 1-for-1 chip swaps) are not "inefficient failures" to maximize utility; they are vital signals of relational labor that establish trustworthiness, index reciprocity, and build long-term social capital. By optimization-forcing the proposals, the Delegate modality strips this relational buffer.

Conclusion. LLMs are increasingly deployed in collaborative settings. The central question is no longer whether they can match human performance, but how the design of human–AI interaction shapes outcomes for individuals and groups alike. We demonstrate that the interaction structure is a first-order determinant of welfare. In a withinparticipants randomized experiment (N= 243), participants preferred the Advisor, yet achieved the highest payoffs when assigned access to the autonomous Delegate. Adjusting for voluntary non-compliance, delegating to the AI yields suggestive individual welfare gains, roughly 1.5× the intent-to-treat estimate. A mechanism analysis reveals that the Delegate advantage arises not from a different AI capability, but from the absence of human intervention. In the Advisor and Coach modes, users modify, override, or ignore AI recommendations, filtering them through a relational lens and pulling the trade distribution back toward human social baselines.

Together, these findings imply that interfaces are not merely a user experience layer but part of the mechanism itself. The bottleneck for AI-assisted welfare in strategic settings is not model capability but the interaction structure through which that capability is delivered.