fMRI and structural equation modeling (SEM) were used to study effective connectivity inside the working memory network in patients at the earliest stage of multiple sclerosis (MS), while performing paced auditory serial addition test (PASAT), a sensitive task to reveal subtle cognitive impairments related to working memory and information speed processing. The path model used for SEM included bilateral connections between left and right BA 46, left and right BA 40, left and right anterior cingulate cortex (ACC), left BA 44 and left BA 40, right BA 44 and right BA 40, and unidirectional ipsilateral connections from BA 46 to BA 44, from ACC to BA 46, and from ACC to BA 44. Experimental data from the two groups fit accurately the working memory model, in patients [chi20(2) = 13, P = 0.877] as well as in controls [chi20(2) = 13.54, P = 0.853]. The omnibus test indicated a significant difference of model fits in patients and in controls [chi40(2) = 160.07, P < 0.0001]. Connectivity strengths from right BA 46 to left BA 46, from left ACC to left BA 46 were lower in patients than in controls, and higher from right ACC to right BA 46, from left to right and from right to left ACC (stacked model). Effective connectivity inside the working memory network appears altered in patients at the earliest stage of MS. Modulation of effective connectivity is present in patients inside the executive subsystems of working memory, and could be related to adaptive cognitive control processes that may limit the clinical manifestation of MS.