Shall deep learning be the mandatory future of document analysis problems?

As the use of deep methods become widespread in the scientific community, causing major changes in systems architecture and position in terms of knowledge acquisition, we report here our insights about how document analysis systems are built. Where does the expertise really lie? In the features, in the decision making step, in the system design, in the data illustrating the problem to be solved? The examination of the practices of researchers in this field, and their evolution, allows us to conclude that the tools that are used, and related issues, have become more and more complex over time. Nevertheless, human skill is needed to activate these tools and to imagine new ones.