Trucey: AI-Mediated Negotiation Coaching

• Trucey is a theory-driven AI system that integrates negotiation and behavioral psychology frameworks to reduce negotiation anxiety and enhance strategic competence.
• It employs a modular pipeline including scenario assignment, personality calibration, few-shot prompting, and role-based rehearsal to deliver personalized coaching.
• Empirical studies reveal that while Trucey effectively reduces negotiation fear, its adaptive guidance shows mixed results in boosting overall negotiation preparedness across personality profiles.

Trucey is a theory-driven, AI-mediated coaching system designed to enhance workplace negotiation readiness, with a focus on addressing both the psychological and strategic barriers commonly faced in high-stakes professional interactions. Grounded in established negotiation and behavioral psychology frameworks, Trucey combines principled, scenario-specific guidance with adaptive, role-based rehearsal. Through iterative design and empirical evaluation across multiple published studies, its architecture reveals both the promise and limitation of conversational AI for complex, recursive preparatory domains such as negotiation. Trucey has also contributed to foundational insights about the interaction between user personality, cognitive load, and the effectiveness of coaching interventions (Duddu et al., 26 Sep 2025, Duddu et al., 1 Apr 2026, Duddu et al., 20 Jun 2026).

1. Theoretical Underpinnings

Trucey's design draws from organizational psychology and negotiation research, integrating:

• Brett et al.’s (2016) 6-Element Negotiation Framework: Structuring negotiation on axes such as strategy development (basic to innovative), management of information asymmetry, exploration of interests (creative, mutual, evaluative), outcomes analysis, long-term relationship goals, and power dynamics (Duddu et al., 26 Sep 2025, Duddu et al., 1 Apr 2026).
• Bradley & Campbell’s Difficult Conversations Phasing: Segmenting negotiation into preparation, in-conversation management, and post-interaction reflection (Duddu et al., 26 Sep 2025, Duddu et al., 1 Apr 2026).
• Behavioral-Psychological Constructs: Grounded in Bandura’s Social Cognitive Theory (emphasizing self-efficacy), Self-Determination Theory (autonomy, competence, relatedness), and Spreitzer’s Psychological Empowerment (competence, meaning, impact, self-determination) (Duddu et al., 1 Apr 2026).

These theoretical pillars are operationalized through staged, context-specific guidance and simulation, aiming to both reduce affective barriers (e.g., fear) and increase cognitive empowerment (e.g., strategic competence).

2. System Architecture and Implementation

Trucey exhibits a modular pipeline integrating personalization and simulation, implemented as a web-based application (Python, Streamlit, GPT-4o-mini API):

1. Scenario Assignment: Users select or are assigned specific workplace negotiation scenarios (e.g., raise, promotion, shift flexibility) (Duddu et al., 26 Sep 2025).
2. Personality Calibration: Users rate their supervisor using the BFI-10, yielding vector embeddings mapped to prototypical supervisor profiles via cosine similarity (Duddu et al., 1 Apr 2026).
3. Few-Shot Prompting and Contextual Layering: The system generates theory-driven advice using prompts incorporating negotiation framework elements at varied “difficulty” or integrativity levels, chunked into micro-lessons (Duddu et al., 26 Sep 2025, Duddu et al., 1 Apr 2026).
4. Role-Based Rehearsal: Trucey simulates the supervisor, adapting pushback tone, language formality, and challenge intensity to the profiled personality, with user feedback collected bi-directionally (Duddu et al., 1 Apr 2026).
5. Iterative Feedback Loop: Users rate realism, challenge, and supportiveness, with these ratings dynamically modulating subsequent guidance and dialogue style (Duddu et al., 1 Apr 2026).

The workflow is illustrated below:

Stage	Key Operation	Underlying Principle
Scenario Assignment	User receives tailored negotiation context	Situational specificity
Personality Calibration	BFI-10 mapping, embedding, profile matching	Contextual personalization
Content Delivery	Few-shot, chunked micro-lessons	Cognitive load minimization (chunking)
Role-Based Rehearsal	Dynamic, theoretically grounded simulation	Exposure reduction, strategic rehearsal
Feedback Integration	User-rated adjustment after every two turns	Adaptive, user-aligned presentation

This pipeline reflects the core assumption that strategically scaffolded, context-personalized dialogue lowers negotiation anxiety and builds competence, though system evaluations reveal several caveats (Duddu et al., 26 Sep 2025, Duddu et al., 20 Jun 2026).

3. Personality Profiling and Differential Effects

Trucey incorporates a personality clustering mechanism rooted in the Big Five Inventory (BFI-10) and ARC typology (Duddu et al., 1 Apr 2026):

• Clustering: Raw trait scores are Z-standardized and input to k-means, with $k=3$ yielding an optimal silhouette score (≈ 0.197). The resulting clusters are:
  • Undercontrolled: High neuroticism, low extraversion/agreeableness/conscientiousness, near-zero openness.
  • Resilient: Low neuroticism, high extraversion/agreeableness/conscientiousness/openness.
  • Overcontrolled: Moderate neuroticism, lower extraversion/openness, mid-range agreeableness/conscientiousness.

Strict outcome analyses show:

• Undercontrolled participants display minimal improvement in self-efficacy or empowerment across any coaching modality.
• Resilient participants achieve the greatest psychological gains from static, self-paced handbooks, with interactive AI sometimes decreasing perceived meaning or autonomy.
• Overcontrolled participants uniquely benefit from Trucey's structured, delivered guidance but experience a tradeoff in linguistic fluency and cognitive complexity.

This personality-sensitive variation establishes a “readiness floor” for AI-driven coaching: not all users are equally positioned to benefit, and pre-intervention regulatory support may be required for vulnerable segments (Duddu et al., 1 Apr 2026).

4. Experimental Methods and Evaluation Metrics

Three studies evaluated Trucey in large-scale, between-subjects experiments (N=267) augmented by semi-structured interviews (Duddu et al., 26 Sep 2025, Duddu et al., 1 Apr 2026, Duddu et al., 20 Jun 2026):

• Conditions: Trucey (theory-driven AI), Control-AI (generic chat AI), Control-Theory (static handbook).
• Measures:
  • Occupational Self-Efficacy (OSS-6): ΔOSS = OSS_post – OSS_pre
  • Psychological Empowerment (PEU): Four subscales, ΔPEU_* = PEU_post – PEU_pre
  • Negotiation Fear/Willingness: Pre- and post-task scales
  • Usability (UMUX) / Appropriateness (IAM)
• Statistical Analyses: Kruskal–Wallis H-tests, Dunn’s multiple comparison tests with Bonferroni correction, t-tests, Cohen’s d effect sizes, embedding-based "framework adoption" via cosine similarity.

Key empirical findings:

• Trucey consistently reduced negotiation-related fear more than generic chat AI (d = –0.27, p < 0.05), supporting exposure effects (Duddu et al., 26 Sep 2025, Duddu et al., 20 Jun 2026).
• Static handbooks yielded higher psychological empowerment (d = –0.40, p < 0.01) and usability (d = –0.32, p < 0.05) than either AI-based intervention.
• No significant difference in actual negotiation preparedness across conditions; willingness to initiate negotiation increased marginally in all arms.
• Lexico-semantic and readability analysis reveal a “fluency tax” for overcontrolled users under AI scaffolding, supporting the need for diagrammatic or alternate scaffolds in certain clusters (Duddu et al., 1 Apr 2026).

5. Conversational AI: Assumptions, Breakdowns, and Cognitive Load

Trucey’s performance illuminated several design assumptions:

• Articulation ≠ Clarification: The presumption that prompting context articulation produces deeper understanding was challenged; users retrieved surface details with little transformative insight, and lacked the means for recursive refinement (Duddu et al., 20 Jun 2026).
• Personalization ≠ Strategic Gain: While context-driven simulation reduced affective fear, it did not measurably enhance users’ strategic competence or empowerment.
• Chunking and Linear Scaffolding Increase Navigation Cost: Incremental delivery of negotiation principles, beneficial in sequential domains, proved suboptimal for recursive, exploration-driven tasks. Users cited increased extraneous cognitive load due to lack of a navigable structural overview (Duddu et al., 20 Jun 2026).
• Ownership and Informational Autonomy: Static materials empowered resilient users through skimmability, bookmarkability, and free recursion across content, features not supported by the enforced linearity of conversational interfaces (Duddu et al., 26 Sep 2025, Duddu et al., 20 Jun 2026).

6. Framework Adoption and Design Principles

Trucey’s prompting structure reliably shifted user output toward integrative negotiation tactics, evidenced by significant increases in structured, innovative strategies and mutual interest exploration, alongside reduced reliance on basic or authority-based approaches (effect sizes d ≈ 1.2–4.0, p < 0.05 for 17/18 Brett elements) (Duddu et al., 1 Apr 2026).

Synthesizing the experimental and qualitative outcomes, the studies propose revised principles for adaptive coaching design:

• Sequencing: Map before Path, Path before Simulation: Prioritize a persistent, navigable overview (Coaching Map) before dialogic deep dives, and confine simulation to user-selected domains (Duddu et al., 20 Jun 2026).
• Outcome-Specific Scaffolding: Competence (self-efficacy) and meaning (value reflection) compete for user resources; fast-tracked handbook access enhances meaning, while stepwise simulation scaffolds competence (Duddu et al., 1 Apr 2026).
• Modular/Hybrid Architectures: Optimal coaching combines static, hyperlinked theory repositories with AI-driven practice sessions, allowing toggling between deep dives and quick-reference modes (Duddu et al., 26 Sep 2025, Duddu et al., 20 Jun 2026).

7. Implications for AI-Mediated Negotiation Coaching

• Personality-Adaptivity: Effective deployment mandates user-specific routing—fast-track for resilient users, modular scaffolding for overcontrolled, pre-regulatory interventions for undercontrolled profiles (Duddu et al., 1 Apr 2026).
• Cognitive Load and Accessibility: Under conditions of anxiety, users demand bullet points, search, and simple navigation—implicating Flesch-Kincaid and similar metrics in the design of output (Duddu et al., 26 Sep 2025).
• Equity Considerations: Undercontrolled profiles overlap with workplace resource deficits, requiring integration with human coaching or robust emotion-regulatory modules to avoid exacerbating inequalities (Duddu et al., 1 Apr 2026).
• Transfer and Authenticity: Users value authenticity and adversarial realism in simulations but question transferability absent physiological cues and unpredictability (Duddu et al., 26 Sep 2025).

A plausible implication is that future AI-mediated coaching systems should foreground user control, transparency of process, and dynamic scaffold calibration, avoiding one-size-fits-all conversational paradigms.

---

References:

(Duddu et al., 26 Sep 2025, Duddu et al., 1 Apr 2026, Duddu et al., 20 Jun 2026)