How to Choose: Building an AI Model from Scratch vs. Using a Pre-Trained Model

Carmen Tovera

May '25 • General Discussion 💬

If you're building an AI-driven product or automating a process, one of the first questions you'll face is:

“Should we build the AI model from scratch, or use a pre-trained one?”

Here’s a quick breakdown to help you (or your dev team) make the right decision

Building from Scratch

When to do it:

You have a very unique problem or data type
You need total control over the model's architecture, behavior, or outputs
You have massive labeled datasets and a solid MLOps pipeline
You’re solving proprietary or regulated problems (e.g., medical, legal, finance)

Costs & Risks:

Requires large computing resources
Takes weeks or months to train
High maintenance overhead

Using a Pre-Trained Model

When to do it:

You’re working with standard problems (e.g., text generation, classification, summarization)
You want to go live fast and iterate
You’re using APIs like GPT-4, Claude, or Hugging Face models
You can fine-tune or prompt-engineer the model for your domain

Bonus: You can fine-tune a pre-trained model with just hundreds or thousands of your examples to make it feel like it was built just for you.

Real Example:

At SalonsMarket, we use pre-trained LLMs for lead qualification and content generation because they’re fast to deploy and highly flexible. But if we were building a pricing engine based on proprietary historical data, we might consider training that model from scratch.

Key Takeaway:

Use pre-trained models for speed and flexibility. Go custom when accuracy, control, or uniqueness demands it.

Let your AI strategy match your business reality.

1 comment