This thesis investigates whether integrating tacit knowledge into Large Language Model (LLM)-based meeting summaries can enrich organizational memory. Following a Design Science Research (DSR) approach, a prototype was developed to transform meeting audio into tacit-enriched knowledge outputs. Initial experiments explored audio-only Speech Emotion Recognition (SER) but revealed significant reliability issues in realistic, multilingual and interruption-rich environments, particularly Swiss German. Consequently, the solution pivoted toward a text-based LLM framework using structured, role-based prompt engineering to extract tacit insights from transcribed meetings.

The artefact generated two summaries, one capturing explicit and a further one to capture implicit emotional and tacit cues. Its effectiveness was evaluated through participant feedback from real meeting contexts, assessing perceived accuracy, usefulness, and potential user acceptance. Results indicate that tacit-enriched summaries provide more nuanced and contextually rich representations than conventional minutes, although users expressed caution regarding over-interpretation and potential surveillance effects.

The study contributes to knowledge management and artificial Intelligence (AI) research by demonstrating the viability of semantic inference as a practical alternative to audio-based emotion detection for tacit knowledge capture. It further highlights opportunities for evolving toward longitudinal, team-based memory systems that support deeper organizational learning and reflective practice.