The emerging field of metabolomics enables researchers to measure concentrations of large numbers of metabolites in biofluids, and to interpret them in connection with the underlying metabolic network, which poses a significant challenge for manual analysis. Given a set of observations on metabolite concentration changes, our goal in this study is to employ automated reasoning, and provide researchers with possible metabolic action scenarios that may have occurred in the body to produce the observed metabolite changes. Our proposed methodology, called the Observed Metabolite Analysis, is to (1) computationally chase the implications of a given a set of metabolite concentration change observations in body fluids, relative to a control subject, (2) eliminate metabolic action scenarios, called hypothesis, that are invalid (i.e., those scenarios that could not have happened) (e.g., increased protein turnover), and (3) rank possibly valid metabolic action scenarios on the basis of pre-defined flux ratio information. We computationally evaluate the proposed methodology with typical metabolomics data, and demonstrate that (a) through consistency analysis against a small number of measured metabolite concentration changes, over 90% of the automatically generated hypotheses are invalidated with no manual analysis, (b) using summarization techniques, the entire hypothesis set is represented by a much smaller (2% of the original) hypothesis set, and (c) performing hypothesis generation and consistency checking in an interleaved manner leads to over 95% improvement in running time.