Four different feed protein evaluation systems were compared on the basis of 157 treatment means obtained from 25 productions trials. Comparisons were made between the British metabolizable protein (ARC, 1984; AFRC, 1992), the French PDI (Verite et al., 1979; Verite and Peyraud, 1989) and the Nordic AAT-PBV systems (Madsen, 1985). Protein values for feedstuffs were calculated using effective protein degradability (EPD) estimated by the nylon bag method and to estimate energy supply for rumen microbial protein synthesis either feed table values or determined diet digestibilities were used. Protein supply (S), requirement (R), SIR ratio and utilization of absorbed amino acids available for milk protein production were calculated using treatment means. The accuracy of the systems to predict differences in dietary protein value was studied by one-way analysis of variance using trial as a factor. Coefficient of variation (C.V.) in SIR ratio and utilization of amino acids for milk protein production were smallest with the AAT-PBV system indicating that this system was the most accurate in predicting differences in dietary protein value. Accuracy of the AAT-PBV system was further improved by reducing the original passage rate of 0.05 h(-1) to 0.02 h(-1) for forages, 0.03 h(-1) for energy supplement and 0.04 h(-1) for protein supplements in calculating EPD values. Estimating microbial protein synthesis from digestible carbohydrates (DCHOs) and rumen degradable protein (RDP) rather than from DCHO alone also improved accuracy. A value of 179 g per kg DCHO + RDP was used to estimate microbial protein synthesis. Since 1995, a modified AAT-PBV system has replaced digestible crude protein for feed evaluation in Finland. Regression analyses showed that the utilization of AAT for milk protein production improved with increasing production level. In feeding recommendations this was taken into account by reducing AAT requirement per kg of corrected milk with increasing production level. (C) 1998 Elsevier Science B.V.