The study of adaptive instructional systems benefits from models that simulate complex cognitive and creative tasks. We present a model that simulates the improvisational techniques used by musicians when interpreting figured bass. This notation combines a bass line with numerical figures. During the Baroque period and still today musicians playing chordal instruments were expected to improvise the accompaniment based on these figures. The task involves a close interplay of perceptual-motor and central cognitive processes. The task is mostly rule-based, lending itself well to be modelled with the ACT-R cognitive architecture. The model incorporates rules for constructing single chords and for voice leading, partly in procedural and partly in declarative format. A third class of rules, pertaining to sequences, is implemented rudimentarily. The model can retrieve complete chords from memory, which occurs only after sufficient modelled experience. In a study we presented a set of exercises to the model and to seven student musicians in two sessions. Thus, we could study the student’s performance and progress measured through harmonic realization, timing, and errors. These indicators are also provided by the model, allowing a comparison. We found that the time predictions of the model are in a realistic range, matching the timing of the less-skilled students. There were also systematic differences between humans and model that shed light on the underlying strategies. However, we demonstrate that the model’s strategies produce appropriate solutions in reasonable time. This qualifies the model to be used as a basis for an adaptive training system to be constructed for teaching and learning to improvise figured bass.