When AI Joins the Table: How Large Language Models Transform Negotiations
Description
This study investigates how Large Language Models (LLMs) transform business negotiations. Through an experiment with 120 senior executives, we examined negotiations with symmetric and asymmetric AI assistance. When only one side had access to LLMs, they gained substantial advantages-buyers achieved 48.2% better deals and sellers 40.6% better outcomes compared to their counterparts. However, symmetric LLM access yielded even more striking results, with 84.4% higher joint gains compared to non-assisted negotiations. This improvement came with increased information sharing (+28.7%), creative solution development (+58.5%), and value creation (+45.3%). Notably, when both sides used LLMs, they relied less on traditional trustbuilding approaches while maintaining fairness, with minimal gain differences between parties (2.2%). Based on these findings, we introduce 'technological equilibrium' to explain how equal AI access transforms negotiation dynamics. While early adopters showed clear advantages, our results suggest that symmetric access ultimately promotes both value creation and procedural fairness through technological parity, enabling integrative outcomes even when trust is limited.
Files
Steps to reproduce
Research Design and Procedure This study used a 2×2 factorial experimental design to analyze the impact of role (Buyer/Seller) and LLM collaboration (with/without) on negotiation outcomes. Each condition was tested 15 times, following the recommendation of Druckman (2005). Participants Our study focused on experienced negotiators to examine LLM effects beyond novice-level enhancement. The study's participant pool comprised 120 senior corporate executives enrolled in executive education programs at leading international business schools and universities. Negotiation Context and Design A modified version of the "Law Library" negotiation scenario provided the experimental context, involving two law firms negotiating the sale of specialized legal resources. This scenario was selected for its potential for integrative and creative options, as it contains multiple issues that can be combined in various ways to create value for both parties. The case involved a law firm splitting into two entities, negotiating with another law firm regarding the sale of specialized law books. The experiment was conducted within Advanced Negotiation courses at leading business schools between May and November 2024. Building on the capabilities of LLMs discussed in our literature review, the training phase focused on applying these capabilities specifically to negotiation contexts. Participants attended a standardized 2.5-hour training focused on applying these capabilities to negotiation contexts. The training comprised: 1. Strategic application of LLM capabilities to negotiation tasks (45 minutes) 2. Negotiation-specific prompting strategies (45 minutes) 3. Hands-on practice with negotiation scenarios (60 minutes) The preparation phase consisted of a 60-minute period during which teams reviewed case materials in separate rooms. Teams in LLM conditions were instructed to apply the prompting strategies learned during training to analyze their negotiation positions and develop potential solutions. This arrangement prevented cross-team communication while allowing participants to develop their negotiation strategies independently. During this phase, participants received detailed written instructions regarding their roles and objectives, along with confidential information specific to their assigned positions. The negotiation phase lasted 35 minutes. Teams were allowed one three-minute timeout, aligning with Carnevale and O'Connor's (1993) findings on the strategic use of breaks in negotiation. During these breaks, participants in the LLM condition were permitted to consult their AI tools, maintaining consistency with their preparation phase resources. Sessions were conducted simultaneously in separate rooms to maintain environmental consistency and prevent cross-team communication.