Automated Mediator for Human Negotiation: Pre-Mediation via a Structured LLM Pipeline

Researchers developed an automated mediator using a structured LLM pipeline to support pre-mediation in human negotiations, decomposing the preparation process into specialized modules for dialogue, preference prediction, critique, and summarization. Human-subject experiments show the system achieves outcomes comparable to professional human mediators on self-reported measures while reducing preference-inference errors by 36%, suggesting scalable AI-assisted negotiation preparation is viable.

This research addresses a genuine friction point in dispute resolution: pre-mediation preparation is valuable but inaccessible due to mediator scarcity and cost. The structured pipeline approach represents a thoughtful engineering solution, separating inference, generation, and evaluation tasks rather than attempting to solve the entire problem with a single monolithic prompt. This modular design mirrors how human mediators actually work and enables parallel deployment across all negotiating parties, which is operationally significant for scalability.

The experimental validation is rigorous within its scope. The researchers compared AI-mediated pre-mediation against professional human mediators using controlled multi-issue negotiation scenarios, measuring both subjective outcomes (trust, confidence) and objective performance (preference prediction accuracy). The 36% reduction in RMSE on preference inference is quantitatively meaningful, and follow-up prompt refinement that reduced excessive affirmation from 36.6% to 16.8% demonstrates the system's tuning responsiveness. However, the findings are narrowly bounded—short-term self-reported measures do not capture whether agreements reached actually prove durable or mutually beneficial in practice.

For the broader market, this work validates that LLM systems can handle complex interpersonal coordination tasks with structured design patterns. The implication extends beyond negotiation: dispute resolution, contract drafting, and other lawyer-adjacent services represent high-value domains where AI mediation could unlock access for underserved populations. The scalability advantage over human mediators is economically significant. The limitation is that the system shows performance parity, not superiority, suggesting displacement risk is moderate for professional mediators but adoption barriers are low for cost-sensitive users seeking preparation support.

• →A modular LLM pipeline for automated pre-mediation achieves outcomes comparable to human mediators on trust and agreement confidence metrics.
• →The system reduces preference-inference errors by 36% RMSE compared to human mediator baselines in controlled negotiation scenarios.
• →Structured task decomposition (dialogue, prediction, critique, summarization) outperforms monolithic single-prompt approaches for negotiation support.
• →Parallel deployment across all dispute parties enables scalable access to pre-mediation without the cost and scarcity constraints of human mediators.
• →Short-term preparation outcomes are comparable to human mediators, but long-term agreement durability and mutual benefit require further validation.