One embodiment provides a method, including: receiving information corresponding to a dialog flow between a conversational agent and a user; identifying, from the received information, dialog states, each dialog state corresponding to a possible response by the conversational agent; classifying a set of the identified dialog states as grounded dialog states, wherein the classifying comprises (i) receiving scoring input representing a desirability for each of the grounded dialog states and (ii) generating a score vector; assigning a score vector to each of the identified dialog states other than the grounded dialog states, wherein the assigning a score vector comprises extrapolating the score vector of a grounded dialog state to dialog states that the grounded dialog state is dependent upon; and computing an aggregate score vector for the dialog flow, wherein the aggregate vector score identifies an overall desirability of the dialog flow.