The instinct is to reach for the largest available model. For open-ended reasoning that instinct is right. For the narrow, repetitive tasks that make up most commercial machine learning, it is usually wrong on every axis that matters.
The tasks in question
Classifying support tickets. Extracting fields from invoices. Deciding which of nine teams a request belongs to. Flagging transactions for review. These have narrow input distributions and a fixed set of correct answers.
Why smaller wins
- Accuracy. A model tuned on ten thousand of your own labelled examples learns your edge cases. A general model knows the world but not your business.
- Cost. Two orders of magnitude is common, which changes what you can afford to run on every record rather than a sample.
- Latency. Small models run in tens of milliseconds, which puts them inside interactive loops.
- Control. You can host it, version it, and reason about why the output changed.
The real cost
Labelled data. Nobody escapes it. The organisations that succeed here are the ones that started capturing decisions as structured outcomes before they needed them. If you are not doing that yet, start now; the model is the easy part.