Assessment and application of Artificial Intelligence to support business processes in a manufacturing SME: The case of BDF Digital

This thesis investigates how BDF Digital can introduce artificial intelligence into internal information-management processes through a controlled and testable pilot. Four candidate application Scopes were initially mapped and compared in terms of information volume, data format, organisational involvement and implementation feasibility. Workplace health and safety was selected because it combined a meaningful organisational perimeter with a comparatively manageable documentary base. The corresponding repository was consolidated from approximately 3,183 to 2,408 files and processed through SCANFILE, a Python-based pipeline developed to read heterogeneous formats, apply OCR where necessary and generate file-level metadata locally with Llama 3.1 8B. The pipeline successfully processed 2,386 files (99.1% of the reference corpus), while manual review showed that free-form category generation required substantial harmonisation and that generated descriptions were considerably more reusable. The prepared dataset was then used to develop a cloud-hosted document-grounded AI prototype providing document retrieval, controlled source selection, question answering and page-level citations. Testing showed that the main limitations concerned retrieval quality, metadata governance, repository maintenance and context management rather than the final answering stage. A subsequent prototype introduced several related improvements, including page-based context control and more flexible document selection. Under the assumptions adopted in the economic analysis, recurring service costs are small relative to the approximately €20,000 initial development investment; the economic return therefore depends primarily on the operational time savings that future use can generate. Overall, the project shows that effective AI adoption in this context depends not only on model capability, but also on repository quality, governance, user training, human validation and continuous maintenance of the underlying information base.