Atmospheric inversions provide a way forward to constrain e.g. surface CO2 fluxes at the regional to global scales using observed atmospheric CO2 concentration and an atmospheric transport model to relate the CO2 concentrations to the fluxes. The atmospheric transport acts as the essential link between the observations and the quantity of interest, in this case the surface fluxes; thus, it is important to quantify the uncertainties associated with the atmospheric transport, or we may overconstrain the solution. Here we present recent results from a new regional-scale ensemble-based data assimilation system with the aim at constraining CO2 fluxes over North America. Through observing system simulation experiments, we explore the predictability of CO2 transport in the atmosphere, and how transport errors impact the posterior CO2 fluxes and their error covariance structure. The ultimate goal is to understand how uncertainties in the atmospheric transport limits our ability to constrain the surface CO2 fluxes, both in the practical sense given our current observing network, and in more ideal cases with near-perfect knowledge of the CO2 concentration state.