The design and empirically evaluation of an LLM-assisted pipeline that classifies heterogeneous aircraft handover documents and extracts metadata from them. The pipeline combines rule-based skimming for initial classification with LLM-based extraction and validation, complemented by a Human-in-the-Loop mechanism that captures expert corrections as structured feedback. The solution targets at least 80% document-level classification accuracy and extraction completeness, a target value derived from comparable LLM-based approaches. By increasing automation, the pipeline is expected to reduce the manual intervention rate and processing time per document compared to the current manual process. The secondary contribution consists of design insights on which architectural decisions, prompting strategies and validation mechanisms prove most effective for applying LLMs to heterogeneous, non-standardized documents in such industrial settings.