I proposed a deep learning solution for a tabular data problem last year.

My colleague suggested XGBoost.

XGBoost won. By a meaningful margin. Trained in minutes instead of hours. Required no GPU. Was more interpretable. Used a fraction of the data.

This happens more often than the deep learning hype would have you believe.

For tabular data — and that's what most business data IS — gradient boosted trees (XGBoost, LightGBM, CatBoost) are still king in 2026. Not always, but more often than not. They're faster to train, need less data, are more interpretable, and don't require expensive hardware.

Deep learning shines on images, text, audio, and sequences. For structured business data in rows and columns? Often overkill.

The best ML engineer uses the right tool, not the fanciest one. Sometimes that's a 175B parameter transformer. Sometimes it's a gradient boosted tree you can train on your laptop during lunch.

There's no shame in simple solutions that work. In fact, there's significant wisdom in reaching for them first.

The measure of engineering skill isn't complexity. It's effectiveness.