A Study on the Design of an AICC System Based on Agentic AI : N/A = Agentic AI 기반 AICC 시스템 설계 연구

다국어 초록 (Multilingual Abstract)

A Study on the Design of an AICC System Based on Agentic AI Yeonghoon Jeong Department of Medical Artificial Intelligence, Graduate School, Eulji University (Supervised by Professor Min Soo Kang, Ph. D.) In modern business environments, contact centers have evolved beyond simple customer support channels, and they now function as core platforms for managing customer experience (CX) and shaping brand trust. However, existing AI Contact Centers (AICC) that rely on rule-based systems or single large language models (LLMs) reveal structural limitations in handling complex decision-making and context-dependent interactions. Data security vulnerabilities, cost inefficiency, and hallucination occurrences have been identified as major obstacles that hinder real-world deployment. To address these challenges, this study proposes a next-generation AICC architecture based on Agentic AI. The proposed system integrates four core intelligence components, which are autonomy, memory, tool-use capability, and reflective learning, in order to build a reflective intelligence framework in which the AI can reason, plan, reflect, and improve without external intervention. A hybrid configuration that combines on-premisess small LLMs (sLLMs) with cloud- based large LLMs is adopted to ensure real-time responsiveness, security, and scalability. The research methodology consisted of literature analysis, requirements extraction, system design, implementation, and validation. We first analyzed the structural and technical bottlenecks of conventional AICC systems to derive key requirements, and these requirements were used to design LangGraph-based multi-agent architecture. Functional agents such as the Planner, Executor, Evaluator, and Policy Agent were structurally integrated, and a testbed that replicates customer service environments in the telecommunications domain was constructed. System performance was evaluated using operational efficiency metrics such as AHT and FCR, policy compliance accuracy, and self-improvement capability, which was assessed through Reflection Gain. Experiments based on representative consultation scenarios, including plan changes, billing errors, and retention consultations, demonstrated that the proposed architecture provides higher response accuracy, better compliance performance, and improved agent–human collaboration efficiency compared to conventional LLM-based AICC systems. The system also exhibited autonomous quality improvement through the reflection loop. Quantitatively, the Realtime API-based speech-to-speech pipeline reduced response latency by 47.4 percent, and the hybrid routing and caching strategies reduced operational costs by 68.5 percent. Sensitive information leakage was fully eliminated, and the hallucination rate decreased from 14.5 percent to 0.4 percent through the combined use of RAG and multi-layered safety mechanisms. This study provides academic and practical value as one of the first attempts to shift the AICC paradigm from automation to autonomy. From an academic perspective, it operationalizes the concept of Agentic AI within an industrial architecture and demonstrates the feasibility of autonomous systems by integrating the Re-Act loop, which consists of reasoning and acting, and the self-reflection mechanism into the consultation workflow. From a practical perspective, it presents a technical roadmap for designing safe and efficient hybrid LLM-based customer service systems that satisfy multiple conflicting objectives such as low latency, cost reduction, enhanced security, and minimized hallucination risk. Furthermore, by identifying future integration opportunities involving affective intelligence, continual learning, and AI governance, this research establishes an important foundation for the development of a human-centered AI consultation ecosystem in which the system can think, reflect, and improve continuously. Keywords: Artificial Intelligence Contact Center (AICC), Public Sector Digital Transformation, Operational Efficiency, Smart Tolling System, Automated Civil Service

목차 (Table of Contents)

• ABSTRACT = ⅰ
• TABLE OF CONTENTS = ⅱ
• LIST OF TABLES = ⅲ
• LIST OF FIGURES = iv
• Ⅰ. INTRODUCTION = 1
• 1. Research Background and Motivation = 1
• 2. Research Objectives and Significance = 7
• 3. Scope of Research = 10
• 4. Expected Contribution = 11
• Ⅱ. Related Work = 13
• 1. NLP/NLU-based Dialogue Systems = 14
• 2. LLM-based Dialogue Systems = 17
• 3. Concepts and Architectures of Agentic AI = 21
• 4. Applicability of Agentic AI to AICC = 24
• 5. Differentiation from Prior Studies = 26
• Ⅲ. Proposed Method = 28
• 1. Problem Definition and Requirements Analysis = 29
• 1.1 Analysis of Limitations in Existing AICC Systems = 29
• 1.2 Definition of Requirements for Agentic AICC = 33
• 2. Agentic AICC Architecture Design = 37
• 2.1 Design Philosophy and Principles = 38
• 2.2 System Architecture and Technology Stack = 41
• 2.3 Multi-Agent Collaboration Framework = 46
• 2.4 Key Technical Components = 49
• 2.5 Memory Mechanism in Agentic AI = 58
• 2.6 Multi-Agent Role Model = 61
• 2.7 Security and Policy Compliance Structure = 64
• 2.8 Performance Optimization Strategy = 65
• 2.9 System Integration and Service Workflow = 66
• Ⅳ. Experiments = 71
• 1. Implementation Environment = 72
• 2. Data Collection and Preprocessing = 74
• 3. Implementation Details = 75
• 4. Experimental Setup and Evaluation Metrics = 76
• Ⅴ. Results = 77
• 1. Latency Optimization Analysis = 78
• 2. Security Enhancement Assessment = 84
• 3. Cost Efficiency Analysis = 86
• 4. Hallucination Mitigation Performance = 88
• 5. Discussion = 90
• Ⅵ. Conclusion = 94
• REFERENCES = 96
• 국문요지 = 121
• LIST OF TABLE
• Table 1. Performance and Quality Indicators of Rule-Based AICC
• Table 2. Representative Core Studies on LLMs
• Table 3. Four Core Components of Agentic AI
• Table 4. Agentic AI Agent Structure
• Table 5. Limitations of Conventional AICC and Agentic AI Improvement
• Methods
• Table 6. Operational Constraints of LLM-based AICC
• Table 7. Agentic AICC System Pseudocode
• Table 8. Technology Stack by Layer in Agentic AICC
• Table 9. Comparison of Realtime API Method and Traditional STT–LLM–TTS
• Method
• Table 10. Comparative Analysis of Self-Refine and Reflection
• Table 11. Comparison of Agentic AI Memory Frameworks
• Table 12. Core Memory Capabilities of Agentic AI and Technical Approaches
• Table 13. Comparison of Multi-Agent Role Models
• Table 14. Comparison of Multi-Agent Planning Models
• Table 15. Detailed Comparison Between the STT–LLM–TTS Pipeline and the OpenAI Realtime API