Although heralded as one of the strengths in AI, raw data supply to artificial intelligence for training has proven to be a mixed bag of successes and failures. While its attractiveness stems from the simplicity, where a minimum a priori inference on the data has to be done, the risk is that during training, the AI machinery focuses on non-relevant collateral patterns, which undiscovered biases in the training data can drive. There are some infamous examples of failures in health diagnosis and image recognition. An alternative is to abandon the idea of feeding raw data, and instead, identify and extract relevant variables from it that can be fed to the AI engine as the sole source of training or used as part of the training information. Furthermore, this preliminary stage can also be subject to a non-supervised process. While this approach is not new, it still lacks general frameworks robust enough to be used in a wealth of areas with minimum specialisation. In this talk, we will present a framework, developed by our group, that has been used in various contexts and has proven its robust nature and effective performance. The framework is based on information theory. It will be explained, and examples will be given in health applications, neuroscience, and dynamical theory