On one hand, finite element analyses have nowadays become a standard in order to design complex soil-structure interaction works, especially when dealing with sensitive structures in the vicinity of the project. On the other hand, reliability design and full probabilistic analysis are more and more used in practice. In this contribution we aim to show that day-to-day 3D finite element probabilistic analyses can now be performed in a reasonable amount of time, using adequate meta-models like Polynomial Chaos Expansions (PCE) or Polynomial Chaos Kriging (PCK). We illustrate this on a study for a tunnel which will be constructed in Lausanne, Switzerland, for a new underground line, still under development. A 3D finite element mesh is constructed, including real stratigraphy, excavation steps, support and bolting. Probability density functions are defined for soil parameters, loads... A sensitivity analysis is conducted first in order to select the relevant input parameters, and then a reliability analysis is performed, the quantity of interest being the settlement of a sensitive building due to the tunnel excavation. Special attention is put in this paper on the selection of the most adequate meta-model, on the error measurement (generalization of the leave-one out error) and on the size of the experimental design needed to construct a reliable surrogate model.