Hybrid optimisation methods using machine learning tools are a hot topic in combinatorial optimization. The AI promise is to learn from past solutions or in real time what are the markers of good solutions and then to guide the resolution of the problem. However, for many hard problems such as VRP, a wide range of powerful solvers have already proven their efficiency whereas the most recent successes in machine learning (e.g deep learning) need a huge amount of training data before reaching a satisfactory performance level. In this paper, we study the efficiency of a features-guided multiple-neighborhood search (FG-MNS) for realistic instances of HFVRP. The solver is based on two steps. In the learning step, a powerful solver (RADOS) is used to generate a set of solutions, which are characterized by a set of features described in previous work (Lucas et al., 2019, Lucas et. al, 2020). Then, a decision tree is built on this training set to determine where are the promising areas in the features spaces. In a second step, called exploitation step, the solver uses some rules extracted from the decision tree to guide the solution toward a promising area. We present a wide range of experimentations with different variants of the FG-MNS solver (offline, online). While the results are promising for a better understanding of what makes a good solution, the real benefits of machine learning in an industrial implementation are questionable and discussed in this work.