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

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

Problem Statement
As Large Language Models (LLMs) transition from passive tools to autonomous agents, a critical gap exists in understanding how different allocations of agency affect outcomes in multi-party, strategically interdependent environments. While prior research has extensively studied human-AI dyads in tasks with verifiable ground truths (e.g., medical diagnosis), there is scarce empirical data on how AI assistance functions in bargaining scenarios where one participant's actions alter the payoff landscape for others. Specifically, it remains unclear whether the "preference-performance misalignment" observed in individual tasks—where users prefer sub-optimal agents that require less cognitive load—persists in group settings, and how different interaction modalities (Advisor, Coach, Delegate) influence collective welfare, adoption rates, and market spillovers.

Methodology
The authors conducted a randomized, within-subjects online behavioral experiment ($N=243$ participants forming 81 groups of three). Participants engaged in three successive multi-turn bargaining games (chip trading) with induced private valuations, a setting known to exhibit human inefficiencies due to risk aversion and fairness norms.

The study operationalized the spectrum of AI agency into three distinct modalities, all powered by the same underlying LLM (Gemini-2.5-Flash) to ensure constant model capability:

1. Advisor: The AI proactively recommends an action (offer or response) with a rationale; the user retains veto power and can modify the suggestion before submission.
2. Coach: The user drafts an action first; the AI provides reactive feedback on the plan, which the user can incorporate or ignore.
3. Delegate: The AI autonomously generates and executes the action on the user's behalf. The user cannot veto or modify the decision, though they can view the rationale.

The experiment utilized a counterbalanced design where each group experienced all three modalities in randomized order. The study measured group and individual surplus (scaled against a Pareto-efficient benchmark), adoption rates (voluntary usage), and user preferences via pre- and post-game surveys. The analysis employed Linear Mixed-Effects Models (LMM) to account for the nested structure of the data, calculating both Intent-to-Treat (ITT) effects and Voluntary non-compliance-adjusted (VNCA) estimates.

Key Results
The study documents a significant preference–performance misalignment and identifies a human filter mechanism as the driver of welfare differences:

• Preference vs. Performance: Participants strongly preferred the Advisor modality (44% of choices) over the Delegate (19%), citing a desire for control and trust. However, groups achieved the highest collective surplus only under the Delegate modality. Neither the Advisor nor the Coach conditions produced statistically significant welfare gains over the human-only baseline at the group level.
• The Human Filter: Mechanism analysis revealed that while AI-generated proposals consistently created higher joint surplus than manual proposals across all conditions, human intervention in the Advisor and Coach modes diluted this benefit. Users frequently modified, reduced, or ignored AI suggestions, reverting to conservative, human-baseline trade patterns (e.g., 1-for-1 swaps) to align with social norms. In contrast, the Delegate mode bypassed this filtering step entirely, allowing the AI's superior optimization capabilities to reach the market.
• Market-Making Spillovers: The Delegate modality generated positive externalities. Unassisted counterparts in Delegate games achieved a 21.6% increase in individual surplus compared to the baseline. The autonomous agents acted as cooperative "market-makers," introducing high-quality proposals that structurally upgraded the shared offer pool, rather than exploiting unassisted players.
• Adoption Frictions: Despite the Delegate's performance advantage, adoption was not universal. Adjusting for voluntary non-compliance, the individual welfare gain for delegates was roughly 1.5 times the ITT estimate, suggesting that the psychological cost of ceding control prevents users from realizing the full potential of the technology.

Significance and Claims
The paper argues that the bottleneck for AI-assisted welfare in strategic settings is not model capability, but the interaction structure through which that capability is delivered. The authors claim that:

1. Interaction Design is a Mechanism: Assistance modalities should be designed as mechanisms with endogenous participation. The "human filter" in advisory and coaching roles acts as a socio-technical translator that prioritizes relational alignment and control over raw material efficiency, often at the cost of collective welfare.
2. The "Inverted Collaboration Paradox": Unlike traditional adoption dilemmas where individuals bear costs for group benefits, delegation here offers a potential win-win (individual and group gains) but is blocked by psychological aversion to relinquishing agency.
3. Design Implications: To realize welfare gains, systems must address the control–delegation tradeoff. The authors propose design paths such as "delegation with veto windows" (allowing autonomous execution with a review period), progressive confidence disclosure to calibrate trust, and group-level evaluation frameworks that prioritize collective surplus over individual satisfaction metrics.

The study concludes that while autonomous delegation yields the highest efficiency, its adoption in real-world multi-party systems depends on designing interfaces that mitigate the psychological costs of ceding control without reintroducing the efficiency-killing "human filter."