Robust predictive models should be based on sound science. Therefore the first step in each computational chemistry project should be an analysis of the scientific context. Based on this a scientific model should be designed which reflects the problem to be solved as close as possible.
But we are currently still quite far from being able to predict all chemical and physicochemical phenomena and parameters just from ab initio by applying fundamental equations. In order to answer real-world questions in plausible time we need to fit a number of parameters to known experimental data.
For the parameterization of robust and predictive semi-empirical models we need good experimental data. According to my experience, this is a sparse resource in most cases. When we have collected such data, we can adjust the missing parameters of the model to the available data and maybe by that detect outliers, which need to be analysed.
But despite of having worked in machine learning already in 1987, I am not a fan of skipping the model building and just leaving the data prediction to machine learning of AI.