The majority of real-world probabilistic systems are used by more than one user, thus a utility model must be elicited separately for each new user. Utility elicitation is long and tedious, particularly if the outcome space is large and not decomposable. Most research on utility elicitation focuses on making assumptions about the decomposability of the utility function. Here we make no assumptions about the decomposability of the utility function; rather we attempt to cluster a database of existing user utility functions into a small number of prototypical utility functions. Having identified these prototypes, we can then effectively classify a new user’s utility function by asking many fewer and simpler assessments than full utility model elicitation would require.