Since ChatGPT arrived, many clients come to us asking for "an LLM". Sometimes that's exactly right. Often, a classic machine learning model would be cheaper, faster and more accurate. Here's how we decide.
What each one is good at
Classic machine learning (regression, decision trees, gradient boosting, small neural networks) learns patterns from your structured data — numbers and categories in tables. It is brilliant at prediction: sales next month, the chance a customer will leave, whether a transaction looks like fraud.
Large language models learn from huge amounts of text. They understand and produce language: answering questions, summarising, classifying free text, writing drafts.
A quick decision guide
| Your problem | Usually best |
|---|---|
| Predict a number or category from table data | Classic ML |
| Forecast demand, prices or churn | Classic ML |
| Understand emails, reviews, documents | LLM |
| Answer questions from your knowledge base | LLM + retrieval |
| Detect objects in images | Computer vision model |
| Very high volume, low cost per prediction | Classic ML |
| Little or no labelled data, lots of text | LLM |
Cost and speed
A gradient-boosting model can make thousands of predictions per second on a cheap server for almost nothing. An LLM call costs more and takes longer. If you need a prediction on every transaction in real time, classic ML usually wins.
Explainability
When a bank, insurer or regulator asks why the model made a decision, classic models are easier to explain. This matters more each year as AI regulation grows.
The best answer is often both
Many of our projects combine them. An LLM reads incoming customer emails and extracts the key facts; a classic model then predicts which customers are most likely to cancel; a dashboard shows the result. Each tool does what it does best.
Questions to ask before you start
- Is my data mostly tables or mostly text and images?
- How many predictions do I need per day, and what can each cost?
- Do I need to explain individual decisions?
- Do I have labelled examples of the right answer?
Unsure which path fits your data? Send us a short description and we'll recommend an approach before you spend anything on development.