The paper describes the major problems encountered during the development of two automatic services delivered through a telephone platform. Service specification, system design and development, and service evaluation and tuning will be described. A significant number of user interactions, including both speech signals and recognizer outputs, have been collected during the usage of the systems. These corpora have been used for testing/retraining both acoustic and language models. In one of the two services (i.e. the automatic payment of the road tax) both context independent and context dependent acoustic models have been compared. In this service, furthermore, an unsupervised training procedure has been evaluated. The second service, a voice portal for accessing financial information, is mainly based on word spotting. We will discuss the effort needed to both develop and tune the related speech recognition grammars.