Designing Conversational AI for Social Robots in Corporate Contexts: A Case Study on Customizing LLMs through Action Research

This paper presents a case study on customizing Large Language Models (LLMs) for social robots in corporate environments. Over a year-long collaboration with an enterprise, we explored how LLMs can be integrated into multimodal assistants that operate across both embodied robot platforms and flexible digital interfaces. Using three iterative action research cycles, we developed and evaluated a scalable customization framework. We introduce Knowledge Interaction Distillation (KID), a method for generating synthetic training data from simulated user interactions, enabling efficient finetuning of task-specific models. Through participatory design, controlled experiments, and field deployment, we identified core requirements and validated a distributed system architecture that supports multimodal interaction. Our findings offer actionable insights for deploying conversational AI for social robots not as isolated systems, but as part of a broader ecosystem aligned with real-world workflows and organizational constraints.