Abstract
Customer service is an important aspect to maintain a loyal customer base and to improve profitability for business. Automated customer support systems including Artificial Intelligence (AI) powered chatting bots (Chatbots), have emerged as an alternative to human-based customer support to reduce the labor cost and improve efficiency. However, current Chatbots utilizing Natural Language Processing (NLP) and machine learning are largely focused on processing open-end questions which could lead to inaccurate answers that resulted in poor customer experience. In this project, we are proposing an innovative AI customer support system with a model-based program to provide customers with very high accuracy responses. The model will be developed using the domain knowledge from business; that is, the commonly asked questions will be used to initiate model building; then the crowd-sourcing input data will be processed continuously to update FAQs. The system will be implemented in an online platform that allows users to interact with the program that constrain and guide them to obtain the right answers. Also, the program will deliver instant answers while predict consumer intents and behavior patterns. The deliverables of this project are an online platform and will be tested by several local industries, including food stores, clinics, automated vending, and restaurants, which have agreed to be pilot testers. The interdisciplinary team is from computer science and college of business. We anticipate the commercial adoption of the platform at the end of the project. The results will be disseminated at a national conference.
Introduction
Traditional customer service, particularly in facility management (like university maintenance or equipment repair), often suffers from bottlenecks caused by manual request processing. While generic chatbots exist, they frequently struggle with domain-specific accuracy, leading to "hallucinations" or irrelevant answers that frustrate users. This project addresses the efficiency gap in Cal Poly’s facility department by moving away from slow, human-dependent response chains toward an automated system capable of understanding intent, classifying issues, and providing instant, high-accuracy feedback without requiring user training.
Methodology
The team employed a multidisciplinary approach, blending software engineering with UI/UX research. Key technical components included:
- Model-Based Interaction: Unlike purely generative bots, this system uses a model-based program initiated with domain-specific knowledge (FAQs and common service requests).
- Algorithmic Backbone: The backend utilizes Markov models to predict and generate intelligent responses based on user input.
- Topic Modeling: NLP techniques were applied to perform topic modeling, identifying and classifying the "core theme" of a request for backend analysis and routing.
- Hybrid Data Architecture: The system utilizes both relational and non-relational databases to handle structured system data alongside unstructured conversational data.
- QR-Code Integration: To bridge the physical and digital worlds, QR code technology allows users to initiate a context-aware support session instantly from a specific location or piece of equipment.
The project followed a "Minimum Viable Product" (MVP) cycle, using iterative usability studies to refine the interface. The development was supported by endowment funding, primarily allocated to software licenses and contracted services to build the cloud infrastructure.
Results and Discussion
The project focused on primary user data gathered from hundreds of responses via online surveys and face-to-face interviews with Cal Poly stakeholders. Real-time conversation data (crowd-sourced input) was used to continuously update the FAQ models.
Key Findings:
- System Deployment: The team successfully built and deployed a cloud-based application capable of analyzing unstructured data in real-time.
- User Validation: Participants reported high satisfaction, specifically noting the system was "easy to interact with without training" and allowed them to "spend much less time on a regular request."
- Contextual Advantage: By using QR codes, the system gains immediate context about the problem's location or equipment, significantly reducing the discovery phase of support conversations.
- Industry Outreach: The project was presented to over 35 global vendors at the National Automatic Merchandising Association conference, demonstrating its potential for commercial adoption.
Conclusion and Future Work
While the system shows great promise, current limitations include the need for minor application logic modifications when tailoring the system to different industries. Future work will focus on enterprise-level scalability testing and moving from a provisional patent to a full filing for the unique Markov-model-based response algorithm. The interdisciplinary synergy between Computer Science and Business proved vital, showing that facility management principles can be generalized across various sectors with minimal adjustment.
Summary updated