Shinhan Card Builds a Highly Efficient AI Chatbot Using Ontology and Small Language Models

Shinhan Card developed a next-generation AI chatbot in collaboration with AWS GenAIIC to overcome the limitations of existing scenario-based chatbots and respond to customers' complex intents and topic switches.

The core challenge was to accurately classify 51 granular intents arising from the nature of financial consultations, while simultaneously addressing regulatory compliance, cost efficiency, and response latency issues. To this end, they decided to adopt self-hostable sLLMs (small language models) instead of large models.

The key technical innovations are built on three pillars:

• Ontology-based intent classification: Intents are structured into Domain, Action, and Tag to prevent information overload in the sLLM and improve classification precision.
• Distributed Agentic AI: Moving away from a single-LLM structure, role-specialized sLLM agents collaborate to handle complex workflows.
• Autonomous system updates: Through inter-agent dialogue and feedback among AI agents, they built a virtuous cycle in which consultation knowledge and logic are self-verified and continuously refined.