Since large language models became easy to call from an API, a pattern has emerged: teams reach for them regardless of the problem. Sometimes that is right. Often it produces a solution that costs fifty times more, runs a hundred times slower and is less accurate than a method that has existed for thirty years.
The distinction is not complicated once stated.
The core difference
Traditional machine learning learns patterns from your structured data to predict a number or a category. It needs labelled historical examples, it is cheap to run, and it gives consistent answers.
Large language models understand and generate language. They arrive already knowing how language works, so they need few or no examples, and they are expensive per call and non deterministic.
Simplified: if the input is a table, traditional. If the input is text, an LLM. If the output must be a number you will act on financially, traditional.
The decision table
Use traditional machine learning for
- Demand and sales forecasting. Numbers from history. An LLM will produce a plausible number that is not grounded in your data. See demand forecasting for retail.
- Credit scoring and risk. Needs consistency, explainability and often regulatory defensibility. An LLM offers none of the three.
- Churn prediction. Structured behavioural data, clear historical outcomes.
- Fraud detection. Speed, volume and consistency matter. See AI fraud detection.
- Price optimisation and recommendations from transaction history.
- Anomaly detection on sensor or transaction streams.
Use a language model for
- Understanding what a customer message means and routing it.
- Summarising calls, tickets, documents, meetings.
- Extracting structured fields from unstructured documents. Invoices, contracts, CVs. See automating invoice processing.
- Answering questions over your own documents, with retrieval. See the knowledge assistant approach.
- Drafting replies, descriptions, and first versions of anything written.
- Classification with no training data, where you can describe the categories in words. This is genuinely useful and often overlooked.
Use both together for
- Support automation: the LLM understands the question, a traditional model or a database lookup provides the answer.
- Document workflows: the LLM extracts the fields, rules or a model decide what happens next.
- Lead scoring: the LLM reads the enquiry text, a traditional model scores conversion likelihood using it as one feature.
This combination is usually the right architecture for anything involving both text and a decision.
The cost difference, concretely
For a classification task at volume:
- Traditional model: cost is mostly the one off training and a small server. Per prediction, effectively free. Response in milliseconds.
- LLM API: charged per call, every call, forever. Response typically in hundreds of milliseconds to seconds.
At a hundred predictions a day the difference is irrelevant. At a hundred thousand a day it is the difference between a rounding error and a meaningful monthly bill. Our note on what running an AI feature costs covers how this scales.
A practical pattern worth knowing
Use an LLM to create your training data, then train a small traditional model to do the work.
You have fifty thousand unlabelled support tickets. Use the language model to categorise two thousand of them, have a person check a sample, then train a small classifier on those labels. You get LLM quality categorisation at traditional model cost and speed.
This is one of the most useful and least discussed applications of language models in business systems.
The mistakes worth avoiding
- Asking an LLM for a number you will act on. It produces confident, plausible, ungrounded figures.
- Using an LLM where consistency is required. The same input can produce different outputs, which is unacceptable in credit, pricing or compliance decisions.
- Training a custom model when a well written prompt would do. The reverse error, and it wastes months on problems that are solved in an afternoon.
- Fine tuning too early. Better prompting and retrieval solve most cases people reach for fine tuning to fix.
Frequently asked questions
Can an LLM do forecasting?
It can produce a number that reads plausibly. It is not deriving it from your data in the way a forecasting model does. For anything you will order stock or set budgets against, use a traditional method.
Is traditional machine learning obsolete?
No. For structured prediction it remains more accurate, far cheaper, faster and explainable. Language models expanded what is possible with text; they did not replace regression.
Which is cheaper to run?
Traditional, by a wide margin at volume. Language models carry a per call cost indefinitely, which is the cost most pilots fail to model before scaling.
How do we choose for our case?
Look at the input. Table means traditional. Text means a language model. Both means both, with the language model handling understanding and the traditional model handling the decision.
Ezitech builds both kinds of system and will recommend the cheaper one when it is also the better one. See our AI solutions or describe the problem you are trying to solve.