Design Science Research dalam Perancangan Sistem Otomasi Notulen Rapat Berbasis Large Language Model dengan Mekanisme Human-in-the-Loop

Abstract

Meeting documentation is often inconsistently produced, causing organizations to lose track of decision-making processes and weakening follow-up accountability. Prior research on large language model-based meeting summarization automation has generally focused on the technical quality of summaries, without considering human verification or the representation of decision-making processes in the resulting minutes. This research aims to design, implement, and evaluate a meeting minutes automation system architecture incorporating two human verification points, at the transcription and summarization stages, using a Design Science Research approach. The architecture is based on client-side integration, directly integrating speech-to-text and large language model services from the client side without additional server infrastructure, developed through seven design iterations and refinement of summarization instructions from a fixed output structure to a Markdown format preserving the trace of proposals and decisions. Demonstration on real meeting audio lasting approximately ninety minutes showed that automatic audio segmentation successfully addressed file size limitations, with a 44% increase in file size resulting in only a 29% increase in processing time. This research contributes a meeting minutes architecture model adaptable by organizations with limited information technology capacity, while affirming the importance of human verification in using large language models for documents that serve as evidence of organizational decision-making.