An automated semi-industrial system with in-line near-infrared reflectance (NIR) for the characterization of the chemical composition of potatoes was designed and constructed, and its performance was tested. The system consisted of the following subsystems: sample crate manipulator, weighing unit for the gross sample weight, potato washing machines with a washing water recycling system, belt for visual inspection of the potatoes, unit for measuring the underwater weight (UWW), industrial rotary saw blade rasp for pulping the potatoes equipped with a sulfite dosage system for inhibiting enzymatic browning of the pulped potatoes, and industrial NIR system for the measurement of the potato composition. The whole system was controlled and operated by a programmable logic controller and process personal computer system. The system was able to process 12 potato samples per hour. Measurements were done to establish the sample carry-over in the system. The carry-over was proven to be well below the maximum acceptable level of 2%. The UWW values established with the automatic system corresponded very well with the UWW data obtained by manual weighting. The day-to-day reproducibility of the UWW system was tested with golf balls. These balls have about the same specific gravity and size as potatoes. The day-to-day reproducibility coefficient of variation of the UWW unit was 0.4%. As a principle of proof, two tentative partial least squares calibration models, one for the starch concentration and one for the coagulating protein concentration in the potato samples, were calculated, applying leaving one out cross-validation. Both models were very promising. The by NIRpredicted starch concentrations showed to be at least as good or even better than the by UWW-obtained starch concentrations. The average difference between the by NIR-predicted and the chemically measured starch concentration was 0.0±0.3% (w/w). For the coagulating protein concentration, the average difference between the by NIR-predicted and the chemically measured concentration was 0.00±0.06% (w/w). In future years, potatoes of a wider range of varieties, growing locations, and growing seasons have to be added to the present tentative model, in order to get a robust NIR model.