The three ways to customise AI for your business — prompting, retrieval-augmented generation and fine-tuning — what each does, what it costs, and how to choose the right one or combine them.
There are three ways to make a general AI model work for your business, and they operate at different levels. Prompting customises the model's behaviour through instructions and examples in the input. Retrieval-augmented generation (RAG) customises the model's knowledge by retrieving your relevant documents at query time. Fine-tuning customises the model itself by training it on your data to adjust its weights. The right approach — often a combination — depends on whether your need is better instructions, access to your knowledge, or consistent specialised behaviour. The sensible default order is prompting first, then RAG, then fine-tuning only where genuinely required, because that sequence runs from cheapest and fastest to most expensive and involved.
These three methods are frequently confused, yet they solve different problems. If the model is capable but you need it to follow your instructions and format, that is a prompting problem. If the model lacks knowledge of your specific, current information, that is a RAG problem. If you need the model to consistently adopt a specialised style, tone or behaviour that instructions alone cannot reliably produce, that is a fine-tuning problem.
Diagnosing which problem you actually have is the key to choosing, because applying the wrong method — fine-tuning to add knowledge that changes weekly, for instance — is costly and ineffective.
Choosing the wrong customisation method wastes money and time. Fine-tuning is the most expensive and least flexible option, yet organisations often reach for it first because it sounds the most substantial, when prompting or RAG would have solved the problem faster and cheaper.
Getting this right has a clear commercial payoff: faster delivery, lower cost, and easier maintenance. Anthropic's 2026 research found data quality and integration to be the top scaling barriers — both are addressed far more cheaply by RAG than by fine-tuning. For Australian mid-market organisations especially, starting with the lightest method that works conserves scarce budget and effort.
Each method operates differently:
| Method | Customises | Cost & effort | Best for |
|---|---|---|---|
| Prompting | Behaviour via input instructions/examples | Lowest | Shaping tone, format, task instructions |
| RAG | Knowledge via retrieved documents | Moderate | Grounding in your specific, changing information |
| Fine-tuning | The model's weights via training | Highest | Consistent specialised style, format or behaviour |
Prompting changes nothing about the model — it shapes each request. RAG leaves the model unchanged but feeds it relevant context at query time, keeping knowledge updateable by editing documents. Fine-tuning actually changes the model, baking patterns into its weights, which is powerful for behaviour but expensive to update and unsuitable for fast-changing facts.
The decisive technical insight: use RAG for knowledge (because it stays current and auditable) and fine-tuning for behaviour (because it bakes in consistent patterns) — not the reverse.
Begin with prompting and a strong system prompt; it resolves more than teams expect and costs almost nothing to iterate. Introduce RAG when the model demonstrably needs your specific or current knowledge. Reach for fine-tuning only when you have a clear, persistent need for specialised behaviour that prompting and RAG cannot meet — and the data to support it.
Edison AI's implementation work follows this escalation deliberately, solving as much as possible at the prompting and RAG layers before considering fine-tuning, which keeps projects cheaper and more maintainable. The methods also combine well: a well-prompted model using RAG, with selective fine-tuning for format, is a common and effective pattern.
Diagnose what you actually need — better instructions, access to knowledge, or consistent specialised behaviour — and match it to prompting, RAG or fine-tuning respectively. Escalate in that order, using the lightest method that works before reaching for a heavier one. Reserve fine-tuning for genuine behavioural needs backed by data, and keep changing knowledge in RAG. Combine the methods where it helps. This discipline delivers customised AI faster and cheaper, and avoids the common, costly error of fine-tuning a problem that prompting or retrieval would have solved.
An AI readiness audit maps the highest-return use cases before you commit to a model or platform.
Prompting customises behaviour through instructions and examples in the input. RAG grounds the model in your knowledge by retrieving relevant documents at query time. Fine-tuning adjusts the model's weights by training it on your data. They customise behaviour, knowledge and the model itself respectively.
Start with prompting — it is fastest and cheapest. Add RAG when the model needs access to your specific, changing knowledge. Consider fine-tuning only when you need consistent style, format or specialised behaviour that prompting and RAG cannot achieve.
Yes, and they often are. A common pattern is a well-prompted model using RAG for current knowledge, with fine-tuning added where consistent format or specialised behaviour is required. They are complementary, not mutually exclusive.
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 vs RAG vs Prompting: Choosing How to Customise AI
