The 1MB limit on datasets is a design limitation of CA process Automation, I think. If you let datasets (via things like attachments) get much larger then the requirement to store every element of the process transaction could quickly cause massive database blowouts.
That being said, I work with partners who have asked what the solution is for things like large attachments, then, with Process Automation.
This thread seems to have a SOAP suggestion though it's not clear whether anyone has implemenetd it. Has anyone else achieved this sort of setup?