Many designers don't like "rules" in a dialogue. They prefer that the design anticipate and respond to natural user behaviors. MI responses to yes-no questions, for example, often seem reasonable to users and "should" be allowed.

If this is your strategy, then the first and most important design feature is to detect unanticipated OOG:

Are you a member of this club?

"I just need some information on your locations."

(ASR incorrectly recognizes MI "yes" response.)

Please say your membership number.

Here, our grammar is big and flexible. That's good when users give us legitimate MI replies—the system is "friendlier" and supports more "natural" user responses. But bigger also means more sensitive to false acceptance. The above dialogue will cause downstream error amplification as the application—now under the false impression that the caller is a member—struggles with subsequent OOG.

Allowing a wider range of user behaviors automatically implies an increased responsibility for anticipating and detecting uncertainty. If you can't or won't do the latter, then don't attempt the former.