Large language models get the headlines, and they've earned them — they're why software can now read an email or draft a response like a person would. But a quieter truth gets lost in the noise: for a large share of real business problems, the best tool is still the kind of machine learning that's been paying its way for twenty years.

Two different kinds of problems

LLMs shine when the input is language: messy emails, documents, conversations, anything where understanding and producing text is the job.

Traditional machine learning — regression, gradient-boosted trees, classic classifiers — shines when the input is structured: rows and columns, histories and amounts, and the job is a prediction or a score.

• "Which of these 4,000 quotes is most likely to convert?" That's a scoring problem. A gradient-boosted model trained on your own history will beat an LLM on accuracy, run in milliseconds, and cost effectively nothing per prediction.
• "How many units will this product sell in March?" Forecasting. Statistical and tree-based methods remain the benchmark.
• "Is this transaction unusual?" Anomaly detection — a solved category long before chatbots arrived.
• "Summarise this complaint and draft a reply." Now you want the LLM.

The criteria that actually decide it

When we're choosing an approach for a client, four questions do most of the work:

1. Is the input text or a table? Tables favour traditional ML, almost always.
2. Does it run constantly? High-volume, always-on predictions make per-call LLM costs add up fast. Small models are effectively free at scale.
3. Do you need to explain it? "The model weights recency and order size most heavily" is an answer you can give a board or a regulator. LLM reasoning is harder to audit.
4. How wrong can you afford to be? On structured prediction tasks, specialised models are simply more accurate. Accuracy compounds.

The honest answer is usually "both"

The systems that work best in practice are rarely one model. An LLM reads the incoming email and extracts the structured details; a small classifier scores the lead; an agent routes the result and drafts the follow-up. Each part does what it's genuinely best at.

Choosing the boring model when the boring model wins isn't conservative. It's engineering.

The skill isn't reaching for the newest tool. It's knowing which decades of work you get to stand on.