August, a Swedish multi-modal spoken dialogue system featuring an animated agent, was used to collect a database of spontaneous computer-directed speech. The system was designed with several simple domains rather than a single complex one. It was installed in a public location in the center of Stockholm and was available for the general public during six months. The people who interacted with the system were given little or no information on what they could expect the system to understand. In this paper, we draw on the experiences gathered in course of the development of the August system and discuss findings in a database of more than 10,000 utterances. The aim of the paper is to address the issue of how the different types of utterances found in the database reflect the demands and strategies of the users. The categorization of the speech input into utterance types and the subsequent analysis of these types will be discussed, and possible implications for modeling future dialogue systems will be covered briefly.