Projects consisting of maintaining nuclear power plants often involve tasks that are documentation- intensive, and must satisfy strict requirements in safety, confidentiality, and traceability. These tasks, that are associated with knowledge work, can be highly time consuming. This thesis inves- tigates how large language models (LLMs) can be implemented to assist in project management workflows from both technical and social perspectives. This was done through a qualitative study of a process within Forsmark nuclear power plant.

The results indicated that interview participants saw value in AI-support for tasks such as assis- tance and drafting in writing reports, evaluating and summarizing documents, coding, translating, and educating. However, there were also fears of implementing AI, as it might cause over-reliance on the technology and can hallucinate. Based on identified requirements an on-premise solution is proposed, centered on RAG with different modes depending on use case. Deployment of such a system is recommended to be incremental to allow for early testing, starting with a semantic search only mode first, as that was identified to potentially be of value. The study concludes that an LLM-integration is possible if designed around the identified requirements, and implemented as socio-technical system, with clear guidelines for use and training for employees.