Fine-tuning requests usually mean one of two things: missing knowledge or wrong style

Fine-tune a model on a document and the document's contents end up in the weights. The dev.to post that sets out these price bands says facts learned that way cannot be updated, audited or cited, and the model has no way to tell you when it is drawing on them incorrectly ^[8]. Retrieval works the other way round. The material is fetched from your own content at request time and placed in the context before the model answers, and the weights are untouched ^[6]. All three techniques change the output, but only retrieval changes what the model knows ^[9].

The post is blunt about the request itself: "It is one of the most expensive sentences in enterprise AI, and it is usually wrong" ^[18]. Its reading is that the sentence covers two different complaints, and that fine-tuning is a poor fix for the first and an expensive fix for the second ^[2].

The £20,000 floor under a fine-tune is four times the £5,000 ceiling on prompt work ^[2]. Run the sequence as far as retrieval and stop there, and the worst case is £65,000, fifteen thousand under the top of the fine-tuning band ^[3]. Everyone budgets for the training compute, and on the invoice it is usually the smallest item ^[12].

The engineering windows invert the prices. Fine-tuning is quoted at two to eight weeks of engineering, retrieval at four to twelve ^[11][12]. Fine-tuning costs more over fewer weeks because the dataset sits outside the engineering estimate and is human work ^[12].

These are the author's stated prices for typical UK delivery on a mid-sized business application ^[13]. For them to transfer, your source content has to be about as messy as whatever set the £15,000 to £60,000 retrieval band, because the post puts most of that build cost in ingestion: getting documents out of the systems that hold them, chunking them, handling permissions so users only retrieve what they may see, and keeping the index current ^[11]. Your domain experts also have to cost, and be as free to write examples, as the ones behind the dataset line ^[12].

Fine-tuning is worth the budget when the target is behaviour that is hard to describe in words and easy to demonstrate: consistent format, tone, task-specific patterns ^[7]. The post argues that working through the options in order costs less than jumping to the end, even when you end up at the end anyway ^[17]. Start with thirty to a hundred real inputs and known-good outputs, because without them no later change can be shown to have helped ^[14]. Then the prompt: be specific about format, give examples of good output inline, state what to do when the answer is not known, which the post says resolves a surprising proportion of complaints about verbosity and structure ^[15]. Then try a larger or different model, an afternoon's evaluation and frequently cheaper than any customisation work ^[16].