In the study of human±machine systems, the need to have a model of the user is by now taken for granted. The model can be used both as support for design and analysis and as a representation of the user that resides somewhere in the machine. When it comes to the practice of modelling, two characteristic approaches can be recognized. The ®rst focuses on the how of modelling, and is concerned mainly with the structure and contents of models. The second focuses on what is being modelled, and is concerned mainly with the functioning or performance of the model. The ®rst approach has dominated human±machine systems research for several decades, and has led to orthodoxy in modelling by which certain structural characteristics are accepted without questioning. This unre¯ective attitude to modelling has been criticised several times, although with little e ect. In taking the second approach and focusing on what should be modelled, two important issues are that human performance varies in level of control, i.e. in terms of how orderly it is, and that thinking and acting take timeÐ and occur in a context where time is limited. Although it is clearly essential that user models can account for these characteristics, very few existing models are capable of doing so because they focus on internal information processing rather than on performance in a dynamic environment. The paper describes a type of functional model, called contextual control models, which shows how it is possible to account for both di erent control modes and how performance is a ected by time. Indeed, control and time are intimately linked and loss of one may lead to a loss of the other. The contextual control model distinguishes among four characteristic control modes (strategic, tactical, opportunistic and scrambled) and two time parameters (time to evaluate, time to select) that are seen relative to the available time. Finally, a number of applications of contextual control models are described.