In order that the planner conforms to the current theory of dialogue planning, a planning model must be developed. One of the most popular ways of specifying dialogue plan structure is by a dialogue grammar, consisting of hierarchical plan rules. In a system that does not use a user model, there is no consideration that the acts in the plan might have preconditions that are evaluated in the context of the mental state of the agent. Instead, each plan structure and resulting dialogue outcome is equally permissible, no matter what the beliefs of the agent are. In some obvious circumstances, such a planner will fail to be efficient. For example, a user model might be used to discount a plan in which an agent will not be able to answer a question. This failure would be due to a failed precondition of the informing part of the questioning plan, which needs to be evaluated in the context of the user's beliefs. A system with no user model would have no way of determining that the plan is bound to fail. This model of hierarchical planning with preconditions is in keeping with the plan construction theory of Carberry [10], and therefore it is adopted for the planner. Hierarchical planning is powerful enough that any set of finite-state rules traditionally used in finite state planners has an equivalent set of hierarchical rules [43]. From the above example, it is clear that a user model will improve the performance of the planner, and that its presence coincides with the use of preconditions to dialogue acts. Furthermore, the user model should be nested, since the agent's own beliefs about whether a precondition is satisfied may differ from those of the user, and may differ again from the user's beliefs about the system's beliefs. The plan under consideration will therefore be evaluated differently from the perspective of each agent, resulting in a different choice of contribution to the plan for each perspective [53]. Therefore the system needs to be able to evaluate the plan from as many perspectives as there are levels in the belief model.
Different agents can have different beliefs about plan rules as well as about the domain state. This is less the case with routine rules such as a question and answer pair, but more the case for meta-level plans whose subject is a domain plan that is constructed using different rules. As a result the different agents will have different expectations about the course of the meta-level plan. The planner is required to treat plan rules as subject to belief in the same way as the domain state is. Alternative ways of constructing a plan will then be subject to preconditions about the correctness of the plan rule used. These preconditions can be evaluated in the same way as action preconditions. Often in BDI models such beliefs about plan rules are called "capabilities". The use of plan rule beliefs opens the possibility of planning dialogues between an expert teacher and a novice student, or one between cooperative experts who exchange information about their expertise so that each can form a correct plan. With different beliefs about both plan rules and the domain state, it becomes clear that when the agent takes the perspective of the user (or the perspective of the user's model of the system and so on), each perspective will look upon a different plan structure.