Real-world settings impose constantly changing demands on cognition and behaviour. These task demands can often be predicted by the context, and the implicit learning of these probabilistic context-task associations may enhance task performance. While previous studies have focused on how task cues that are either probabilistic or implicit affect task-switching performance, the present study investigated how people learn and use contextual cues that are both implicit and probabilistic within a cued task-switching design. Participants showed response speed benefits when engaging in tasks that were predicted to be more likely by a preceding contextual cue. However, this probabilistic cueing effect was only seen when specific contextual cues were associated with task probabilities (Experiment 2), and not when contextual categories were associated with task probabilities (Experiment 1). The findings provide support for automatic activation of multiple task sets; a model of multiple concurrent task set activation and representation is proposed. Taken together, our findings suggest that people can implicitly learn probabilistic associations between specific contexts and tasks, and can use information from contexts to guide adaptive behaviour in dynamic environments.