DESIGN AND EVALUATION OF A CONVERSATIONAL AI BOT FOR HELP DESK SUPPORT USING THE RASA FRAMEWORK AND DIET CLASSIFIER

Abstract

The digital era has transformed customer service expectations, with users now demanding immediate, accurate, and continuously available support that traditional human-operated help desks struggle to deliver. This study presents the design, development, and evaluation of a Conversational Artificial Intelligence (AI) bot for help desk support, built on the open-source RASA machine learning framework and powered by the Dual Intent and Entity Transformer (DIETClassifier) for joint intent classification and entity recognition. The methodology combined stakeholder-driven requirements gathering, agile design and development, and iterative testing. The bot was trained for 100 epochs on a curated dataset of help desk interactions dominated by Windows operating system support queries, covering tasks such as system updates, boot and network troubleshooting, malware protection, password resets, file recovery, and disk and driver management. Results show that the transformer-based DIETClassifier accurately recognised and classified a wide range of user inputs, from simple greetings to elaborate technical questions, while the UnexpecTEDIntentPolicy provided robustness against atypical or out-of-scope queries. Initial testing demonstrated a considerable improvement in response quality over prior rule-based systems, with the bot maintaining conversational context across multiple turns. Interaction logs, however, revealed a degree of repetitiveness in generated responses, indicating scope for data refinement and more diverse response selection. Overall, the study demonstrates that a lightweight, framework-based conversational AI can substantially improve help desk efficiency, availability, and operational cost, while highlighting emotional intelligence, escalation to human agents, and data privacy as key areas for future enhancement.