“We want an AI model of our own.” We hear this often, but it usually means three different things with very different costs and timelines. This guide helps you work out exactly what you need.
Four ways to “customise” AI
| Approach | What it solves | Cost and time | When? |
|---|---|---|---|
| Better prompts and instructions | Format, tone and answer quality | Very low; days | Always start here |
| RAG | Answering from your documents and data | Medium; weeks | When the model must know your knowledge |
| Fine-tuning | Fixed, repeatable behaviour, format or style | Higher; needs quality examples | When prompting and RAG weren’t enough |
| Training from scratch or classical ML | Forecasting and scoring on structured data; training a language model from scratch is rare | High | Only for special needs with plenty of data |
- The model doesn’t know something
Like prices or a technical document: RAG.
- The model doesn’t behave right
Like output format or tone: fine-tuning.
- Your knowledge changes often
RAG; you only update the knowledge base.
The recommended order: simple to heavy
OpenAI’s official guidance on improving accuracy suggests the same path: start with a prompt, add fixed examples, then a retrieval step (RAG), and finally, if needed, fine-tuning. These methods can be combined. For example, one assistant can be fine-tuned and also answer from your documents.
RAG or fine-tuning? A simple rule
- If the problem is “the model doesn’t know something” (prices, policies, technical documents): RAG.
- If the problem is “the model doesn’t behave correctly” (output format, tone, fixed sorting): fine-tuning.
- If your knowledge changes often: RAG, because you only update the knowledge base and don’t retrain the model.
- If you have both problems: combine them.
What data do you need?
For RAG, your existing documents are a good start, as long as they are organised and current. For fine-tuning you need correct input-and-output examples. OpenAI’s guidance suggests starting with at least about 50 high-quality examples. Quality and variety matter more than the raw number.
How do you measure accuracy?
- Collect 30 to 100 real questions or tasks from your daily work, including hard cases.
- Define the correct answer or pass rule for each.
- Test and compare every approach on this same set.
- Sort the errors: not knowing, making things up, wrong tone.
- Before launch, review the responsible-AI checklist.
What if the model must also take actions?
When a model must work with tools and move through several steps, you enter the world of AI agents, where guardrails and access control matter even more.
Before deciding, check what a chatbot and AI model cost so your choice fits your budget.
Reduce the risk of building a model with a proof of concept before payment.
For a real example of AI running on the device itself, see offline face-recognition attendance: how Hazeri works.
Frequently asked questions
Should we train a language model from scratch?
Almost never for an ordinary business. The cost and data needed are huge, and prompting, RAG or fine-tuning an existing model usually reaches the same result.
Can RAG and fine-tuning be used together?
Yes, and it’s common: fine-tuning fixes behaviour and format, while RAG supplies current, company-specific knowledge.
Can you run the model on our own server?
If data privacy matters, open-source models can run on your server. The choice between an outside model and your own depends on accuracy needs, cost and how sensitive the data is.
How do we know the model works well?
With a test set of your own real tasks. Compare results before and after each change and sort the errors.