In the XACML impl I worked on it the thought had been that the policy modeller would be part of the info security team, even in a PAP this takes some specialized skill. However without a PAP it fell to developers to craft the XML policies. There was also a need to delegate control over some of the details to a variety of different people (business users, customer power users). They wouldn't write policies per say but have control over some of the "constraints" defined by the policies and the UI would be simplified greatly to where it needed little to no training/additional skill.
I like your thinking on the DSL.. that idea had been bouncing around in my head as well. Although, I was thinking it could directly drive the PDP... I see the value in XACML being the data and logical model and not so much in the fact that it uses XML. That is to say I don't see any particular value in generating XML other than the fact that the current PDP drives off of it. Many of the COTS drive off a database representation of the entities. Also, I think a DSL may be a good interim step in creating a usable PAP, but that ultimately it needs a UI on top of it for easiest learning curve.