The instinct to fine-tune is usually a translation of "we have proprietary data and we want the model to know it." Fine-tuning is one way to express that. It is rarely the best way.
Retrieval is the best way when (a) your data changes more often than once a quarter, (b) you need citations or auditability, or (c) you don't have the volume of clean, labeled examples that fine-tuning rewards. That's almost everyone.
Fine-tuning earns its keep when the task is a stable transformation: structured output formatting, classification with a fixed taxonomy, a tone of voice that retrieval can't enforce. The rest of the time, retrieval with a good chunking and reranking strategy will beat a fine-tune at a fraction of the lifecycle cost.