AI Models: Pretrained vs Fine-Tuned

Pretrained vs fine-tuned: what’s the difference?

The customisation spectrum

There’s more than just “pretrained” and “fine-tuned.” Think of it as a spectrum:

The AI model customisation spectrum — from zero effort to maximum investment

Feature	Effort	Cost	When to use
Pretrained (as-is)	None — use the model directly	Lowest (pay per token)	General tasks: drafting, summarising, brainstorming, translation
Prompt engineering	Low — craft better instructions	Same as pretrained	When you need specific output formats, tone, or behaviour without changing the model
RAG (grounding)	Medium — connect to your data	Moderate (retrieval + inference)	When the model needs to answer questions from YOUR documents, not just general knowledge
Fine-tuning	High — retrain on your dataset	Significant (compute + data prep)	When you need domain-specific language, consistent style, or specialised knowledge baked into the model
Train from scratch	Very high — build a new model	Extremely high	Almost never for business — only if you're a model provider or have unique requirements no existing model serves

💡 Exam tip: The customisation decision tree

The exam expects you to know WHEN to recommend each level:

1. Start with pretrained + prompt engineering — this solves 80% of business needs
2. Add RAG when the model needs YOUR data (not general knowledge)
3. Fine-tune only when prompt engineering and RAG aren’t enough — e.g. the model needs to consistently use your industry terminology or match a very specific output style
4. Never recommend training from scratch for a business scenario — it’s almost always wrong on the exam

Business tradeoffs: what leaders need to know

As a business leader, the pretrained vs fine-tuned decision comes down to five factors:

Factor	Pretrained Model	Fine-Tuned Model
Cost	Lower — pay only for inference	Higher — compute for training + ongoing inference
Time to value	Fast — hours to days	Slow — weeks to months
Data requirements	None (or minimal via RAG)	Requires curated training dataset
Maintenance	Provider updates the model	You must retrain when data or requirements change
IP and differentiation	Same model everyone else uses	Your competitive advantage — trained on your unique data

Scenario: Ravi’s build decision

Ravi (CTO, TechVantage) is building an AI assistant for developers. He evaluates:

• Option 1: Pretrained GPT-4o with prompt engineering — Fast, cheap, works well for general coding questions. But it doesn’t know TechVantage’s internal APIs or coding standards.

• Option 2: GPT-4o with RAG — Connect the model to TechVantage’s internal documentation. The model retrieves relevant docs before answering. Good balance of customisation and cost.

• Option 3: Fine-tuned model — Train a model on TechVantage’s codebase and documentation. More consistent output, uses internal terminology naturally. But expensive and needs regular retraining.

Ravi’s decision: Start with Option 2 (RAG). It gives 90% of the benefit at 20% of the cost of fine-tuning. Fine-tune later only if RAG isn’t sufficient.

💡 Why 'start with RAG' is usually the right answer

For the exam and for real life: RAG is often the sweet spot. It lets you ground a pretrained model in YOUR data without the cost and complexity of fine-tuning. Most organisations should start with RAG and only fine-tune if they can demonstrate a clear gap that RAG can’t close.

Model types you should know

Model	Type	Key Characteristic	Business Use
GPT-4o	Large language model (general)	Highly capable, multimodal, expensive	Complex reasoning, content creation, analysis
GPT-4o mini	Smaller, faster LLM	80% capability at 20% cost	High-volume, latency-sensitive tasks
Phi (Microsoft)	Small language model	Runs locally, low cost, open-weight	On-device AI, privacy-sensitive scenarios
Llama (Meta)	Open-source LLM	Free to use, customisable	Self-hosted deployments, research
DALL-E	Image generation model	Creates images from text descriptions	Marketing visuals, concept art, prototyping
Whisper	Speech-to-text model	Transcribes audio in 90+ languages	Meeting transcription, accessibility

Key flashcards

Question

What is the difference between a pretrained and a fine-tuned model?

Click or press Enter to reveal answer

Answer

A pretrained model has broad, general knowledge from training on massive datasets. A fine-tuned model starts with a pretrained base and is additionally trained on domain-specific data to improve performance on particular tasks.

Click to flip back

Question

Where does RAG fit on the customisation spectrum?

Click or press Enter to reveal answer

Answer

RAG sits between prompt engineering and fine-tuning. It grounds a pretrained model in your specific data without changing the model's parameters — often the most cost-effective customisation approach.

Click to flip back

Question

When should a business fine-tune a model instead of using RAG?

Click or press Enter to reveal answer

Answer

Fine-tune when you need the model to consistently use domain-specific terminology, match a very specific style, or when RAG retrieval isn't providing sufficient quality. If RAG solves the problem, fine-tuning is usually unnecessary.

Click to flip back

Question

What is a Small Language Model (SLM) and when would you use one?

Click or press Enter to reveal answer

Answer

An SLM (like Microsoft Phi) is a compact AI model that runs on-device or with minimal compute. Use when you need local processing, low latency, low cost, or data privacy — the model never sends data to the cloud.

Click to flip back

Knowledge check

Knowledge Check

Dr. Anisha Patel is advising a healthcare company that needs an AI assistant to use medical terminology consistently and follow strict clinical documentation standards. What level of model customisation should she recommend?

AUse a pretrained model as-isBUse prompt engineering onlyCUse RAG with medical knowledge basesDFine-tune a model on clinical documentation

Check Answer

Knowledge Check

Elena's firm wants to quickly deploy an AI tool that helps consultants find answers in the company's proposal archive. What's the most cost-effective approach?

AFine-tune a model on all past proposalsBBuild a custom model from scratchCUse a pretrained model with RAG connected to the proposal archiveDTrain a small language model to memorise the proposals

Check Answer

🎬 Video coming soon

Next up: AI Cost Drivers and ROI — understanding tokens, pricing models, and building the business case for AI investment.