Underfitting implies that the predictive model generated during the learning phase is not adapting well to the Training Set.
In other words, the predictive model is incapable of capturing the correlations of the Training Set. Therefore, the error cost in the learning phase remains high and the predictive model will not generalize well on the data it has not seen. Finally, the model will not be viable because the prediction errors will be many. In this case, the model is suffering from Underfitting.