Fine-tuning is worth it later than you think

Fine-tuning is the most-asked and least-often-justified AI investment. In most projects, prompting (with careful few-shot examples), retrieval (hybrid + reranking), and model selection (cheap model for easy queries, frontier for hard ones) cover the quality goal — at a fraction of the cost and time of a training run.

The decision becomes real when two things happen: the eval plateau is above your quality bar even with the best prompt and retrieval, or a smaller tuned model is materially cheaper than a prompted frontier model at equivalent quality. Those are the two honest reasons to fine-tune.

When you do tune, most of the win comes from data quality, not training tricks. Curate, label, and dedupe a dataset you'd be happy to defend. Use LoRA first — cheap, fast, reversible. Measure against the same eval suite as the baseline, including regression checks on base skills the model used to have.

Deploy as a private endpoint with the same infra as the rest of your stack: quota, cost attribution, rollback, version tags. A tuned model is an operational commitment, not a one-time experiment.