Strategically
Model the domain with the experts and end-users (e.g. Eventstorming, DDD, UX). Model "as is" and then "could be", and test/prototyle/mock it with actual users as soon and as often as possible. Automating a paper mess yields an automated mess, and misses an opportunity to learn as a group and improve. Including end users helps ensure you're not building a fantasy/nightmare.
Tactically
Model the workflow for the process in question, noting all information currently collected and how (forms, calls, external systems, etc), then work backwards from the outputs to identify unnecessary information and steps