Large number of antennas in Massive MIMO offer a significant spatial diversity, which makes them an attractive possibility for use in wireless settings that require very high reliability. However, in 5G ultra-reliability is coupled with low latency into ultra-reliable low-latency communications (URLLC). This is very challenging as an efficient use of Massive MIMO depends critically on training, which consumes significant resources when the latency requirement is very tight. In this paper we address this problem by exploiting the sparsity of the propagation channel and therefore rely on estimation of a small number of instantaneous channel coefficients. This leads to robust beamforming and departs from the conventional use of the instantaneous channel state information (CSI) at each transmit antenna. We compare the performance of maximum ratio transmission based on the conventional least-squares estimation of all channel coefficients and the one based on the estimation of the fading coefficients of the channel features i.e. the singular vectors of the covariance matrix. The singular vectors are assumed known and unchangeable over a long term. The results show that this approach makes massive MIMO a feasible technology in URLLC scenarios.