Structural equation modelling was used to study the change of connectivity during a visual task with continuous variation of the attention load. The model was based on areas defined by the haemodynamic responses described elsewhere [Mazoyer, P., Wicker, B. & Fonlupt, P. (2002) A neural network elicited by parametric manipulation of the attention load. Neuroreport, 13, 2331–2334], including occipitotemporal, parietal, temporal and prefrontal (lateral and medial areas) cortices. We have studied stationary- (which does not depend on the attentional load) and attention-related coupling between areas. This allowed the segregation of two subsystems. The first could reflect a system performing the integration step of the visual signal and the second a system participating in response selection. The major finding is the mutual negative influence between the lateral and medial parts of the prefrontal cortex. This negative influence between these two brain regions increased with the attention load. This is interpreted as a modification of the balance between integration and decision processes that are needed for the task to be efficiently completed.