Regression models with ARIMA errors (Reg-ARIMA) are nowadays commonly used in seasonal adjustment to remove the main deterministic effects (outliers, ruptures, calendar effects) from the raw data before decomposing the corrected series into trendcycle, seasonality and irregular. The main seasonal adjustment programs (X-13Arima-Seats, Tramo-Seats, JDemetra+) implement these models in an automatic and very user-friendly way. This facility hides in fact a real complexity and in certain cases a lack of robustness which can escape the user. In the presentation, we draw attention to the real difficulties of implementing these models through concrete cases: the estimation of a leap year effect, the estimation of breaks and even the estimation of an ARIMA model.