Over the last decade, energy harvesting from ambient vibration sources has received substantial attention. Piezoelectric materials are widely used to convert ambient vibrations into electrical energy due to the efficiency of their electromechanical transduction. The finite element method is classically used to model this kind of system. Among its advantages, it makes it possible to capture the influence and effect of the different geometrical, mechanical, and electronic parameters of the problem on its electromechanical response and to study in particular the dependence between the eigenmodes of the system and its frequency response. The amount of electrical energy recovered by the system is largely influenced by the uncertainty of its geometric and physical (electrical and mechanical properties of the piezoelectric layers) parameters. The Monte Carlo method is widely used to assess this influence. Nevertheless, its prohibitive calculation time, in the case of dynamic modeling by finite element method, limits its use. Metamodels are proposed in the literature as a surrogate model allowing a robust modeling of these systems and a considerable reduction of the calculation time for analysis of uncertainty and sensitivity. In this context, this paper focuses on the metamodeling of a piezoelectric energy harvester based on polynomial chaos expansions (PCE) to predict the electrical output with low computational cost. The metamodel is built using a database created from finite element simulations previously fixed by the design of the experiment. The metamodeling is based on a regression technique using six degrees polynomial. The surrogate model is then used for a sensitivity analysis which consists of calculating Sobol indices with negligible computational cost. The main objective is to sort the uncertain parameters in a hierarchical order according to their influence on the harvester’s responses. Simulation results using the metamodel show that the length of the harvester and the thickness of the piezoelectric layer are the most influential parameters on the electrical output, while the width of the harvester and the thickness of the piezoelectric layer are the least influential on the harvester output. These results are shown to agree with those obtained by Monte Carlo simulations using the finite element model.