The labour economics literature on signalling assumes workers know their own abilities. Well-settled experimental evidence contradicts that assumption: in the absence of hard facts, subjects are on average overconfident. First we show that in any equilibrium of any signalling model, overconfidence cannot make players better off. In order to obtain more detailed predictions, we then introduce a specific on-the-job signalling model. We show that at fully-separating equilibrium, overconfident workers choose tasks that are too onerous, fail them, and, dejected by such a failure, settle down for a position inferior to their potential. Such a pattern leads to permanent underemployment of workers, and inefficiency of the economy. For the case of unbiased workers uncertain about their own value, we determine a necessary and sufficient condition for the existence of fully-separating equilibrium.