For conversatonal assistants, language models offer the potential benefit of being able to generate responses to the widest posisble range of queries that adhere to a policy, without the need for a premediated conversational design, which is inherently hard to design optimally for all queries.