This paper aims to show that regularities emerge from the strongly stochastic properties of human behavior. To this aim, the paper presents the application of a new Knowledge Discovery in Database (KDD) process, called Timed Observations Mining for Learning (TOM4L), on the timed data provided by a smart building of offices of the southeast of France during 12 months from April 2011 to March 2012. The TOM4L process produces then 12 behavioral models of the white-collar workers of the office, one for each month of the studied period. This sequence of models put on the light the strongly stochastic properties of human behavior since they differ significantly from one month to another. This illustrates the intrinsic difficulty of discovering behavior rules from the timed data provided by a smart environment. Nevertheless, regularities clearly emerges from this sequence of behavioral models that are closely linked with seasons. Two seasons, a cold season of five month and a warm seasons of seven months, are clearly identified with this sequence but to this aim, more abstract models are required. Finally, this paper shows that the TOM4L approach is clearly operational and powerful for human behavior modeling in smart environments but higher abstraction levels of representation must be defined to discover more general behavior rules.