Recent advances in on-line metal [U, Mn, Cr] reduction systems for the automated analysis of δDVSMOW in water and organic solvents greatly increases sample throughput. However, several sources of uncertainty are inherent in these systems, including analytical reproducibility, fractionation of reference gas in the inlet system, and memory effects, as well as error introduced during data normalization that must be quantified. A careful consideration of these sources of uncertainty suggests that accuracy to <1‰ may be difficult to achieve due to the combined effects of analytical error and uncertainty in the calibration of secondary laboratory standards. However, for internally consistent data sets, it may be possible to measure relative differences between samples at the ∼0.5‰ level or below. In addition to describing methods for data reduction, strategies for minimizing and managing uncertainty are presented along with a recipe for greatly increasing reactor life in Cr-reduction systems. Increased reactor life reduces consumable costs by a factor of 10. Copyright © 2001 John Wiley & Sons, Ltd.