A plain-English definition of fine-tuning — training an existing AI model on your own data to specialise its behaviour — including what it costs and when it is the right choice.
Fine-tuning is the process of training an existing AI model further on your own data to specialise its behaviour — for instance, to adopt a consistent style, tone, format or domain expertise. Unlike prompting (which shapes behaviour through instructions) or retrieval (which supplies knowledge at query time), fine-tuning actually changes the model itself, baking patterns into its weights. It is powerful for consistent specialised behaviour, but it is also the most expensive and least flexible customisation method, and it is frequently reached for when a cheaper approach would do. This entry defines the term and its trade-offs; our fuller guides cover fine-tuning versus RAG versus prompting and RAG versus fine-tuning for the enterprise.
A general AI model is trained to be broadly capable. Fine-tuning takes that model and trains it further on a focused dataset of your examples, nudging its behaviour toward your specific needs. The result is a model that has internalised patterns from your data — a particular writing style, a domain's terminology, a required output format.
The defining characteristic is that fine-tuning modifies the model. This makes it the right tool for behaviour you want consistently and intrinsically, but the wrong tool for knowledge that changes — because anything baked into the weights can only be updated by training again.
Fine-tuning is often misunderstood and misapplied. Organisations reach for it first because it sounds the most substantial form of customisation, when prompting or retrieval would solve their problem faster and far more cheaply. Knowing when fine-tuning is and is not appropriate prevents expensive, avoidable projects.
The correct division is straightforward: use retrieval for knowledge (because it stays current and auditable) and fine-tuning for behaviour (because it bakes in consistent patterns). Anthropic's 2026 research found data quality and integration to be top scaling barriers — both addressed more cheaply by retrieval than by fine-tuning. For Australian organisations watching budgets, understanding fine-tuning's cost and proper use is a direct guard against wasted spend.
Fine-tuning involves:
Efficient methods such as LoRA and QLoRA (parameter-efficient fine-tuning) reduce the cost by adjusting only a small part of the model, and instruction tuning and RLHF are related techniques used by providers. Even so, fine-tuning requires quality data, compute and skill, and it produces a model that must be maintained — a meaningful ongoing commitment.
Before fine-tuning, confirm that prompting and retrieval genuinely cannot meet the need, because they are far cheaper and more flexible. Fine-tuning earns its cost only when you have a clear, persistent need for specialised behaviour and the quality data to train it.
Edison AI's AI training and implementation work helps organisations choose the right customisation method, reserving fine-tuning for genuine behavioural needs and using prompting and retrieval where they suffice. For the decision in full, see our guides on fine-tuning versus RAG versus prompting; the practical insight is that fine-tuning is a precise tool for behaviour, not a default first step.
Understand fine-tuning as a way to specialise a model's behaviour by training it on your data — powerful but the most expensive and least flexible customisation method. Before approving any fine-tuning, confirm that prompting and retrieval cannot achieve the goal, and reserve fine-tuning for consistent specialised behaviour backed by quality data. Keep changing knowledge in retrieval. For the full decision framework, read our guides on fine-tuning versus RAG versus prompting; the practical takeaway is to treat fine-tuning as a deliberate, well-justified choice rather than a reflexive first move.
See how the pieces fit together in a real build on our AI implementation page.
Fine-tuning is training an existing AI model further on your own data to specialise its behaviour — for example to adopt a consistent style, format or domain expertise. It changes the model itself, unlike prompting or retrieval, which leave the model unchanged.
Use fine-tuning when you need consistent specialised behaviour — style, tone, format or domain handling — that prompting and retrieval cannot reliably achieve. It is not the right tool for adding knowledge that changes, which belongs in retrieval.
It is the most resource-intensive customisation method, requiring quality training data, compute and expertise, and re-doing the work when the model updates. For many use cases, prompting or retrieval achieves the goal at far lower cost.
Edison AI helps Australian businesses move from AI curiosity to practical implementation, with workflow design, team training and measurable outcomes. Tell us about your setup and we'll come back with a sequenced plan grounded in the same thinking you just read.
Article: Fine-Tuning: What It Is, What It Costs and When to Use It
