For most predictions, a language model is the wrong tool

Because for numeric prediction it is usually worse, slower and more expensive. Forecasting, ranking, pricing, churn scoring and anomaly detection depend on patterns across millions of rows of structured history — exactly what gradient-boosted trees and similar methods are built to exploit, at a fraction of a cent and a few milliseconds per prediction. Language models earn their place on unstructured input: text, documents, images, conversation. Most real systems use both, each on the part of the problem it suits, and we are explicit about which is doing what.

It depends on the shape of the problem more than a fixed row count. Tabular prediction is often workable with a few thousand well-labelled examples and becomes solid in the tens of thousands. Seasonal forecasting needs at least two or three cycles of history, so a couple of years for anything annual. Rare-event detection depends on the number of positive cases rather than the total — a hundred confirmed frauds among millions of transactions is a harder starting point than the volume suggests. Computer vision can often start small using a pre-trained model and a few hundred labelled images per class.

A focused predictive model on data that is already accessible typically takes eight to twelve weeks from framing to production, and most of that is data preparation and integration rather than modelling. Where the data needs assembling from several systems first, add a month or two. We run a two to three week feasibility assessment before quoting a build, because the honest estimate depends almost entirely on what your data turns out to look like.

Monitoring and a retraining plan. We track the distribution of inputs, the distribution of predictions and, wherever the true outcome eventually arrives, live accuracy — which gives an early signal when behaviour shifts. Retraining runs on a schedule or a drift trigger, each candidate is compared against the model currently in production, and promotion happens only if it wins. Where the environment moves fast, such as pricing or fraud, that cycle is short by design.

For the model types we recommend in regulated settings, yes. Tree-based and linear models support per-decision feature attribution, so you can show which factors moved a specific outcome and by how much. We provide a model card documenting training data, intended use, known limitations and measured performance by segment, retain the inputs and outputs for each decision, and keep a human review step wherever a person can contest the result. If a use case genuinely requires an approach we cannot explain adequately, we will say so before it is built.