Contemporary organisational processes evolve with people’s skills and changing business environments. For instance, process documents vary with respect to their structure and occurrence in the process. Supporting users in such settings requires sophisticated learning mechanisms using a range of inputs overlooked by current dynamic process systems. We argue that analysing a document’s semantics is of uttermost importance to identify the most appropriate activity which should be carried out next. For a system to reliably recommend the next steps suitable for its user, it should consider both the process structure and the involved documents’ semantics. Here we propose a self-learning mechanism which dynamically aggregates a process-based document prediction with a semantic analysis of documents. We present a set of experiments testing the prediction accuracy of the approaches individually then compare them with the aggregated mechanism showing a better accuracy.