In the context of a project aiming to build human-behaving robots for process automation, named entity recognition becomes one of the first tasks we need to solve. In this paper we present our experience on building NER models for recognizing specific entities of interest, with the help of the state-of-the-art pre-trained BERT model. Noticing that the model built with the help of a general knowledge dataset scores poor results in retrieving entities specific to our particular use cases, we constructed two datasets tailored for our working context and trained BERT-based models on it. We show that properly constructing the specific datasets is sufficient in order to obtain a good entity classification performance, without further increasing the model learning time.