To guide the design of advice-offering user-assistance software, "Wizard of Oz" techniques were used to observe the interaction between users of a graphical statistical package and a human playing the role of a simulated intelligent advisory system. The results emphasize the complexities of advisory processes. Video data for 34 cases of advice seeking, giving, and following were analyzed in detail. The evidence indicates that clients followed prescriptive advice effectively and efficiently in slightly more than half the cases. For other cases, clients performed twice as many actions as needed in three times as much time and never reached prescribed states. A hypothesis that observed advice-following difficulties were correlated with advice abstractness was not supported. Rather, it seems advice did not match well with client's knowledge of the system. Impacts on advisory system design are discussed.