Hi,
to me it makes no sense to use the SpamAssassin contrib by michael in a setting like yours. The whole stuff depends on active users, that check their /junkmail folders, and put false positives into hamlearning and spam into junkmailmissed folders.
You can only achieve that with direct access to these folders on the SMEserver. Which to my knowledge is not possible via exchange/outlook.
An autolearn-level of 12.00 does not make sense either, you get all the spam.
If you really want to use bayes filtering in this kind of setup, you should perhaps install assp on your smtp-system. This allows to blacklist, whitelist by mail and trains the wordlists by checking the actual traffic of your users.
Also your users can individually put wrong positives as "notspam" by email by simply sending them to a special account.