research - 2026-07-27

GPT Is Not the Right Tool for Every Data Problem

Why large tabular forecasting tasks often need LightGBM or CatBoost instead of a GPT model, and what to do before choosing an AI approach.

People still reach for GPT when they have a business dataset and want an answer fast. I get why. It feels like the shortest path from a spreadsheet to something that sounds useful. But if the task is forecasting sales, product demand, or another problem built on large tabular data, a GPT model is often the wrong tool. It was built for text and context, not for rows of numeric history.

That is where boosted models like LightGBM or CatBoost usually make more sense. They fit structured columns, seasonality, product history, and the kind of signals that show up in business tables all the time. If the question is “what should we order, and in what quantity,” the model needs to read the table itself, not the transcript of the table.

A mistake I see a lot is people assuming they can send the dataset to ChatGPT and get a real forecast back. GPT can help in smaller ways. It can summarize a few months, write Python code, or give you a first-pass analysis if the data is already reasonably clean. But if the source files are messy, split across formats, or badly structured, the work starts earlier: preprocessing, aggregation, and feature design.

That part matters more than people want it to. If you want a useful prediction, you have to decide what the model should actually learn from the business. Seasonality. ABC analysis. Product attributes. Inventory behavior. Whatever matters in that category. The model is only as good as the inputs you shape for it, and weak inputs usually give you a nice-looking mistake.

We ran into this recently in a Zenit-Auto project on sales prediction and product demand. The goal was simple to say and annoying to do well: predict demand with low error so the business can order the right products in the right quantities. That is not a “just ask GPT” problem. It is a modeling problem, a data problem, and a judgment problem.

We got to a WAPE of 43%. Not magic. Just choosing a model that fit the task, then doing the boring work around it properly. That is usually the real lesson here: before you ask which AI model to use, ask what kind of data problem you actually have.

Planning an AI automation initiative? Read the AI automation POC guide.