All insights
6 min readMachine learning, LLMs, Method
Formula, ML, or LLM? Choosing the right method for every metric
Not every question needs a large language model. The skill is matching the method to the metric and staying provider-agnostic.
There's a temptation right now to answer every question with an LLM. It's powerful, it demos beautifully, and it's frequently the wrong tool.
Three kinds of questions
- Deterministic: aggregations, ratios, statistics. A formula is faster, cheaper and exactly right.
- Pattern-based: forecasting, anomaly detection, scoring. Machine learning earns its keep.
- Language-shaped: summarizing, classifying, reasoning over text. That's where LLMs shine.
The skill is in the routing
A serious analytics system decides, per question, which method to use, and it doesn't marry you to a single AI provider. Models change monthly, and your analytics shouldn't have to.
Use the cheapest method that's correct. Reserve the heavy machinery for the questions that need it.
That discipline is what keeps an AI analytics platform fast, affordable and trustworthy as it scales.