There are three ways to make an AI system know your business: tell it, hand it the documents, or train it. The industry calls these prompting, RAG, and fine-tuning, and businesses reliably reach for them in exactly the wrong order. The expensive one sounds serious, so it gets requested first. Here's the order that respects your budget.

Start by telling it (prompting)

A system prompt is standing instructions: who the assistant is, how it should sound, what rules it follows, what it refuses to do. It's the cheapest lever in AI and the most underrated. A well-written prompt playbook — tone, formats, examples, edge cases — solves the large majority of "the AI doesn't sound like us" complaints for the price of an afternoon.

Reach for it when: the problem is behavior, tone, or format. Which is most of the time.

Then hand it the book (RAG)

When the assistant needs facts — your prices, policies, product specs, past proposals — retrieval is the answer. The system fetches relevant chunks of your documents and answers from them, citing sources. Update the document and the AI is current. No retraining, no waiting.

Reach for it when: the problem is knowledge, especially knowledge that changes. Prices change. Policies change. RAG doesn't care.

Train it last, if ever (fine-tuning)

Fine-tuning adjusts the model itself on hundreds or thousands of your examples. What it's genuinely good at: locking in a very specific output format, a house style at high volume, or domain shorthand a general model fumbles. What it's bad at: storing facts. A fine-tuned model still guesses; it just guesses in your voice. Add the cost of building training data, evaluating results, and redoing it when base models improve, and the bar gets high.

Reach for it when: you're producing one kind of output at real volume, prompting has demonstrably hit its ceiling, and you have clean examples to train on. All three, not one.

The cheat sheet

  • "Make it sound like us" → prompt.
  • "It doesn't know our stuff" → RAG.
  • "It knows our stuff but formats it wrong every time, at scale" → consider fine-tuning.
  • Not sure → prompt first. It's an afternoon, not a project.

The combinations stack, and most production systems use two of the three: a strong prompt plus retrieval covers nearly everything a small business needs. The principle underneath is older than AI: buy the cheapest thing that works, and prove it doesn't before upgrading.