Abstract:

This article investigates data-sovereign deployment of large language models (LLMs) in small and medium-sized enterprises (SMEs) and similar organizations. External cloud-based LLM services raise concerns regarding confidentiality, data sovereignty, data control, and regulatory or contractual compliance, as LLMs typically process data in unencrypted form. The article therefore motivates rethinking such LLM deployments by examining the local, on-premises use of open-weight generative AI models in the hardware-constrained environments of SMEs as an alternative. It proposes a design-science-oriented research agenda based on prototypes, feasibility studies, case studies, and open-source reference implementations to develop a catalog of application patterns for data-sovereign LLM deployment. Application patterns are positioned between low-level software design patterns and reference architectures, capturing and structuring reusable solutions for recurring technical, socio-technical, and regulatory challenges