Computed tomography (CT) scanning of logs makes it possible to appearance grade virtual sawn timber before the log is sawn. The data consisted of a knot structure on the surfaces of the virtual sawn timber from the CT scanning. The knot structure was graded as being suitable or not for a specific product by the existing rule-based approach, and used to create a set of descriptive statistical variables used by two machine learning models. The PLS models were trained on two quality references; the quality grade of the finished product, or the image-grade based on images of the sawn timber, extracted from the dry-sorting stations automatic grading system and graded by two experienced researchers. The results show that the two PLS models perform equally well when sorting sawn timber to the customer, indicating that the quality references are equally useful for training a PLS model. The PLS models both delivered 93% of the dried sawn timber to the customer, leaving very little sawn timber with customer-specific properties at the sawmill, of which 89% and 90% of the delivered sawn timber passed the intended product’s quality demands. The rule-based approach delivered 85% dried sawn timber with a 73% pass rate.