In the risk management context, the extreme-value methodology consists in estimating extreme quantiles-one hundred years return period or more-from an extreme-value distribution adjusted on data. In this communication, we quantify the extrapolation limits associated with extreme quantile estimations. To this end, we focus on the framework of the block maxima method and we study the behaviour of the relative approximation error of a quantile estimator dedicated to the Gumbel attraction domain. We give necessary and sufficient conditions for the error to converge towards zero and we provide a first order approximation of the latter. We show that extrapolations can be greatly limited depending on the data distribution.