CELIA [Redmond 1992] is a multi-strategy learner that, among other things, acquires cases through observing an expert. With all other learning turned off, CELIA improves its predictions of the expert’s actions dramatically. However, performance does not monotonically increase with more cases since 1) some problems are harder than others, 2) an approprlate case still may not be available, 3) there is the possibility of retrieving the "wrong" cases. We report results from experiments with two different problem sets, both from the domain of automobile diagnosis. We discuss the variation in performance at different levels of experience in the two studies.