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

TL;DR

This study investigates how three AI-assisted modalities—Advisor (proactive recommendations), Coach (reactive feedback), and Delegate (autonomous action)—affect welfare, adoption, and externalities in a three-player bargaining game with fixed model capabilities. Using a within-subjects randomization (N=243), the authors find that Delegate access yields the highest group and individual surplus, partly through a market-making effect that improves the trading environment for all players, including non-users. Despite Delegate’s welfare benefits, participants show a strong preference for the Advisor modality, revealing a preference-performance misalignment that highlights adoption barriers in agentic systems. The results emphasize that interface design and endogenous participation are crucial for translating autonomous AI capabilities into tangible welfare gains in social, strategic settings, and they reveal meaningful externalities that emerge when delegation shapes the strategic environment. The work provides an empirical baseline for designing agentic systems in multi-party ecosystems and underscores the need to align user preferences with system-level efficiency through interface and governance choices.

Abstract

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 \textit{access to} a single LLM assistance modality: proactive recommendations from an \textit{Advisor}, reactive feedback from a \textit{Coach}, or autonomous execution by a \textit{Delegate}; all modalities are powered by an underlying LLM that achieves superhuman performance in an all-agent environment. On each turn, participants privately decide whether to act manually or use the AI modality available in that game. Despite preferring the \textit{Advisor} modality, participants achieve the highest mean individual gains with the \textit{Delegate}, demonstrating a preference-performance misalignment. Moreover, delegation generates positive externalities; even non-adopting users in \textit{access-to-delegate} treatment groups benefit by receiving higher-quality offers. Mechanism analysis reveals that the \textit{Delegate} agent acts as a market maker, injecting rational, Pareto-improving proposals that restructure the trading environment. Our research reveals a gap between agent capabilities and realized group welfare. While autonomous agents can exhibit super-human strategic performance, their impact on realized welfare gains can be constrained by interfaces, user perceptions, and adoption barriers. Assistance modalities should be designed as mechanisms with endogenous participation; adoption-compatible interaction rules are a prerequisite to improving human welfare with automated assistance.

Figures (19)

• Figure 1: Overview of the experimental design and contributions. Participants (N=243) engaged in three-person bargaining games with access to three LLM assistance modalities: Advisor (proactive recommendations), Coach (reactive feedback), or Delegate (autonomous actions) --- with within-subject game randomization. We find evidence of a preference-performance misalignment, positive spillover effects in the Delegate games, and heterogeneous outcomes for different classifications of users.
• Figure 2: Relevant game components. Panels 1. and 2. show interface properties visible across all modalities. Panel 3 illustrates the proposal generation interface shown in each of the treatment modalities. More detailed figures of game interfaces are provided in Appendix \ref{['app:dl-interface']}.
• Figure 3: Access to delegation increases both group and individual surplus. Panel (A) shows the comparison of scaled group surplus gain. Panel (B) shows the comparison of scaled surplus gain for individuals.
• Figure 4: Top: Frequency of assistance usage by negotiation round. Bottom: Coefficient of regression predicting AI usage as a function of the negotiation round. Significance levels: $^{***}p<0.001$, $^{**}p<0.01$, $^{*}p<0.05$.
• Figure 5: Receiver surplus distributions of accepted trades. Each panel compares the surplus of AI-assisted sender offers with non-AI-assisted offers. A significant upward shift in receiver surplus is observed only in the Delegate condition ($p < 0.01$), indicating that Delegate agents propose trades that are mutually beneficial.