Models generalize better to standard file system operations (ls, cat, bash, etc.) than to custom-designed tools built for specific business scenarios, because the latter lack sufficient training data to support them.