Named Entity Recognition (NER) is a fundamental task in natural language processing that involves identifying and classifying entities such as names, locations, and organizations within text. Although public NER resources such as annotated datasets and annotation services exist across various domains, no single resource typically supports all entity types required for specific downstream applications. Additionally, the availability of training data to effectively develop NER systems for different domain classification schemas is often limited due to constraints on time, quality, and annotation costs. In this paper, we propose to address these issues by a transfer learning approach, validating the hypothesis that with limited resources, the target domain labels and their distributions can be learned by exploiting features obtained from the source domain. The proposed approach achieves comparable performance with respect to fine-tuning state-of-the-art transformer-based models, using a limited amount of resources in terms of GPU, CPU and RAM.