In the digital era, faced with increasing information overload, organizations are finding it difficult to keep up with the necessary description work. Auto-classification tools that automatically assign metadata are promoted as the solution to the metadata gap, but can they fulfil their promise in concrete, real-life cases? This article reports on a small proof-of-concept project, carried out in an international organization, to investigate the feasibility of using semantic rule-based auto-classification for its official documents. It presents the test batch, the taxonomies used, the organization of the work, the evaluation through a user study and the results of the project. The new task of preparing taxonomies, so that they can be applied directly to natural language, is also considered. Despite some issues resulting from time constraints, the proof of concept proved to be largely successful. It was a valuable opportunity to better understand how auto-classification works, and what the prerequisites are for rolling it out to the whole organization.