In this context, strategy means the type of inference used. This can determine the order in which questions are asked. It may be important, not only to have a strategy that gets the right answer, but one which asks questions in the same kind of order that a human expert would. For example, some inference engines would require that all answers to be input in advance. This may irritate a user, especially if he knows some of these answers may never be used. And for some questions, it can lead to silly situations: ``Your sex please?''; ``Male''; ``Are you pregnant?'' ...But systems that don't ask until they need data can also cause problems. Some sets of rules may result in a system that constantly switches from one topic to another and appears annoyingly unfocussed.
It's also important that the user be given a chance to volunteer data. With a pure backward chaining system, no data can be input until it's requested. If answering one question reminds you of the answers to another few, you should be allowed to volunteer the answers on the spot, rather than having to wait.