When is fine-tuning actually worth it versus just writing better prompts or adding RAG?

This is a question that comes up in practically every serious AI implementation discussion and IBM's fine-tuning explainer https://www.ibm.com/think/topics/fine-tuning is the right starting point for thinking through the decision properly.

The decision framework the article implies: prompting is the fastest and cheapest approach for improving output quality on a task, RAG is the right solution when the problem is missing or outdated knowledge, and fine-tuning is appropriate when the problem is tone, style, format, domain vocabulary or task-specific behaviour that cannot be reliably addressed through prompting or context.

Fine-tuning is not a shortcut to a more capable model. A fine-tuned model on a small dataset for a specific task will outperform a base model on that task and underperform it on general tasks. That is the right trade when you are building a production system with a specific, well-defined use case.

The cost consideration is the practical gating factor most articles gloss over. Fine-tuning requires labelled data, compute time, model hosting, and the ongoing work of evaluating and retraining as requirements change. For many use cases, better prompting or a RAG architecture solves the problem at a fraction of the investment.

Has anyone in the community fine-tuned a model for a production use case? What made it worth the investment over prompt engineering or RAG?