Loss-based seismic design involves targeting selected levels of earthquake-induced loss (e.g., deaths, dollars, downtime). Achieving this in the practice, without resorting to tedious trial-and-error procedures, requires using simplified loss assessment strategies for a set of seed structures, thus selecting the one(s) complying with the selected loss target. As the first step towards practical loss-based design for bridges, this paper investigates the trade-off between simplicity (modelling effort and computational time) and result accuracy in the seismic loss analysis of reinforced concrete (RC) bridges. The study involves a set of case-study bridges imagined in a high-seismicity site. For each of them, a benchmark estimation of the expected annual loss is obtained running non-linear time-history analyses of refined multi-degree of freedom models. Such estimates are obtained with alternative, simplified strategies including analytical non-linear static analyses (i.e., pseudo pushover) using the displacement-based assessment paradigm, coupled with the capacity spectrum method; the same non-linear static analyses coupled with a surrogate probabilistic seismic demand model based on a Gaussian process regression. This sensitivity analysis also tests the impact of the assumed functional form for the probabilistic seismic demand model (power-law or a bilinear). Although the results are preliminary and potentially subjected to refinements, the use of the displacement-based pseudo pushover together with the capacity spectrum method shows the potential to be included in a practice-oriented loss-based design paradigm for bridges. Conversely, adopting a surrogate model-based loss-assessment would first require challenging some of its simplified assumptions.