Financial institutions continue to rely on decades-old mainframe systems written in COBOL to support critical operations. As expert developers retire and modernization projects accelerate, banks face a dual challenge: preserving functional knowledge while ensuring smooth migration to modern languages such as Java. This paper reports on the design of an orchestrated multi-LLM pipeline that automatically generates functional and technical documentation from COBOL source code within secure, air-gapped environments. The approach combines expert-validated pre-processing, structured prompt engineering, and a hybrid evaluation process mixing LLM-as-a-Judge with human review to assess fidelity, readability, and business relevance. The experience gathered from a large-scale implementation at BNP Paribas shows that while single-LLM configurations maximize factual fidelity, orchestrated multi-LLM setups better interpret and capture business semantics—an essential prerequisite for COBOL-to-Java migration. These observations highlight practical trade-offs between accuracy and abstraction in zero-fine-tuning settings and outline a governance-aligned path for integrating generative AI into regulated financial institutions.