Talking AI — Fine-tuning: the most-pitched promise, found in 9% of production models

Fine-tuning means retraining a model on your own data so it picks up your vocabulary and your way of working. It is the most-pitched and least-practiced technique: in a Menlo Ventures survey of 600 IT decision-makers, only 9% of production models were fine-tuned, against 51% built on RAG, where the model reads your documents at question time instead of memorizing them. The following edition of the same survey, published in December 2025 among 495 decision-makers, confirms that fine-tuning remains niche.

The gap comes down to cost of ownership: a retrained model freezes knowledge that must be retrained at every price, product-line or procedure change, whereas a document base updates within the day. When a vendor offers to train a model on your data, ask one question: what happens when that data changes, and who pays for the retraining? Before signing, have both architectures priced separately on a single scope, and commit to retraining only once the document-based approach has been tested on real cases and found wanting.

Source: Menlo Ventures, “2024: The State of Generative AI in the Enterprise”, 20 November 2024 (600 IT decision-makers, United States); 2025 edition, 9 December 2025 (495 decision-makers).