In all areas of human activity, the generation and storage of data are universal and growing. The extraction of knowledge from this data is a generally complex and very demanding task in terms of human resources, both for domain experts and data analysts. The need to automate and simplify the process, in particular for the domain specialist, is increasingly imperative for the effective use of data in decision-making processes. In this work, based on an extensive survey of literature regarding analytical tasks, we propose a taxonomy of analytical tasks (independent of the domain), and evaluate the performance of several state-of-the-art Large Language Models (LLMs) in the decomposition of questions of the domain into a set of analytical tasks, for 5 different domains. Only after this decomposition we perform the necessary tasks on the domain data and propose a ranking of visualizations that best respond to the question of the domain, using a Visualization Recommendation System (VRS). The two-step separation: (i) decomposition into analytical tasks; (ii) execution of tasks, including the generation of views, distinguishes our approach from most proposals that use LLMs for analytical activities. Another aspect that distinguishes this work is the presentation of an ordered set of views in a declarative way utilizing Vega-Lite as the rendering environment.