Purpose: To evaluate the accuracy of the CEREC CAD/CAM system in reproducing the maximum intercuspal contacts of the casts, which include posterior teeth preparation for a fixed partial denture (FPD). Materials and Methods: Ten pairs of gypsum casts were mounted in articulators in maximum intercuspal position (MIP) to serve as patient simulation models. Tooth #19 was removed from the cast. Occlusal contacts in MIP were identified with articulating paper, and digital impressions of the casts with unprepared teeth and buccal images in MIP were taken. Teeth #18 and #20 were prepared for an FPD, and full-and half-arch digital impressions of the casts with prepared teeth and buccal images from different sides were taken. In each situation, screenshot images of the virtual casts with occlusal contacts were saved as JPEG files. The proportions of congruence of virtual contacts with cast contacts were analyzed by superimposing screenshot images of the virtual casts onto the screenshot images of the casts with the indicated occlusal contacts in a transparent manner using an image-processing program. The data were statistically analyzed with a paired t-test. Results: The highest percentages of virtually indicated contacts identical to the cast contacts were observed in non-prepared full-arch digital impressions. Comparison of full-arch impressions taken before and after tooth preparation showed no difference for congruence even if the buccal image was taken from the contralateral or ipsilateral side (p > 0.05). After tooth preparation, comparing full-and half-arch digital impressions revealed that half-arch impression showed significantly lower percentages of identical contacts (p < 0.05). When comparing the buccal image side, no significant difference was detected between ipsilateral and contralateral images both for non-prepared and prepared casts (p > 0.05). Conclusion: When there is no posterior antagonist contact following tooth preparation for an FPD, taking a full-arch digital impression and designing the restoration on fullarch virtual models can be advocated.