Choosing between classic NLP and LLMs
Language models can do many text tasks with a good prompt, but they are not always the best tool. A fine-tuned encoder such as BERT or a sentence-embedding model with a simple classifier can be faster, cheaper, predictable and easy to host for high-volume labelling, while an LLM is stronger at open-ended or low-data tasks. Cost and speed often decide it, since classifying millions of short messages is a poor use of a large model.
We test both on a sample of your text. Often the answer is a pipeline: rules and small models for the bulk, an LLM for the hard remainder. That keeps latency and cost down and gives you a clear fallback when a hosted API is unavailable. Agreeing label definitions with your subject experts at the start prevents months of retraining on inconsistent examples.
