This paper presents a model and architecture for non-intrusive user interfaces in the interactive digital TV domain. The model is based on two concepts: non-monolithic rendering for content consumption and actions descriptions for user interaction. In the first case, subsets of the multimedia content can be delivered to different rendering components (e.g., video to the TV screen and extra information to a handheld device). In the second case, we differentiate between actions, handlers, and activators. An action is the description of the user intentions, a handler implements that action, and an activator is the user interface of the action. Because we define actions instead of user interfaces, the implementation of the activators can take multiple forms: conventional user interfaces (using gestures or speech) and intelligent interfaces, in which the actions are derived from a set of parameters (e.g., number of people in the room or distance to the TV).