This paper analyses employment transitions and workers’ skills in Brazil using a random sample from the universe of formal labour contracts covering the period from 2003 to 2018. We develop a novel procedure to derive a measure of occupational distance and internationally comparable skill measures from occupations’ task descriptions in the country under analysis based on machine learning and natural language processing methods, but without usual ad hoc classifications. Our findings confirm that workers who use non-routine cognitive skills intensively experience the highest employment growth rates and wages. Their labour market exit risk is relatively low, occupational and sectoral changes are least common and, in the case of occupational switching, non-routine cognitive workers tend to find occupations that are higher-paid and closer in terms of their task content. Against the same characteristics, routine and non-routine manual workers are worse off in the labour market. Overall, there have been signs of routine-biased technological change and employment polarization since the 2014 Brazilian economic crisis.