UX Research Projects
Agentic AI Role-Play for Insurance Sales Training
@Samsung Life Insurance (#1 life insurance company in Korea)
Executive Summary:
In my current role, I'm leading the instructional and UX design for a ground-up "Agentic AI" role-playing platform designed for 35,000+ insurance sales agents. The project focuses on transforming static training into a dynamic, state-based conversational experience that bridges the gap between theory and real-world sales proficiency.
Core Strategic Pillars
1. Human-Centered Design for a Specific Demographic
Designing for an average user profile of 54-year-old female sales agents required a radical focus on Cognitive Load Theory.
-
From-Scratch UX: I led the end-to-end design of the user flow, stripping away technical friction to ensure the interface felt intuitive rather than intimidating.
-
Cognitive Load Optimization: Every interaction was customized to minimize extraneous cognitive load, allowing the aging learner to focus entirely on the nuances of the conversation rather than navigating the tool.
2. Technical Architecture & State Machine Logic
In collaboration with the engineering team, I defined the behavioral logic of the AI using LangGraph.
-
Conversation States: We architected a system where the AI moves through specific pedagogical and sales "states."
-
Dynamic Transitions: defined the logic gates that determine how the AI flows from one state to another—including "fail-state" conditions where the conversation ends if the user fails to meet specific communication benchmarks.
3. Novel Evaluation Metrics: Trust & Objection Scores
To move beyond binary "correct/incorrect" feedback, I co-developed two proprietary metrics used for both user evaluation and system logic:
-
The Trust Score: Measures the user’s ability to build rapport and emotional safety.
-
The Objection Score: Tracks the frequency and intensity of AI resistance based on the user's performance.
-
Functionality: These scores act as real-time logic triggers. If a user’s Trust score remains too low, the LangGraph logic prevents the conversation from advancing, instead triggering a "pedagogical scaffolding" node to provide corrective coaching.
4. PhD Research Integration: AI Pedagogical Agents
I directly applied my Ph.D. research in Cognitive Science in Education to the system’s instructional design by implementing AI-generated Pedagogical Agents (PAs).
-
Evidence-Based Design: Leveraging the Social Agency Theory and "Persona Effect," I utilized realistic, AI-generated agents to increase user motivation and social presence.
-
Learning Effectiveness: By optimally placing these agents within the learning path, we leveraged their ability to model complex behaviors, further reducing cognitive load and making the training more effective for non-technical users.
Projected Impact
-
Scalability: Automated high-fidelity role-play for 35,000+ agents.
-
Training Cost Reduction: Currently projected to reduce enterprise training costs by 30% by replacing manual coaching with automated, data-driven feedback.
-
Retention: Increased training engagement through high-realism AI personas that mirror actual sales scenarios.