Automate the process you want, not the one you have.
A process that has grown by accident usually contains steps that exist because a system could not do something in a previous decade, approvals nobody remembers requesting, and rework caused by a step earlier in the chain. Automating that faithfully makes it permanent. The first task is separating what the process is for from how it currently happens.
The exceptions are the work.
The standard path is normally a small share of the volume and almost all of the demonstration. The value is in what happens to the rest: the incomplete request, the out-of-policy amount, the case that needs a human judgement. An automation that handles only the standard path moves the effort rather than removing it, and often makes the exceptions harder to see.
We map, decide, then build in slices.
Map the process as it runs, with volumes and handling time. Decide per step whether to remove it, simplify it, automate it or leave it alone; not everything should be automated. Design the exception routes explicitly, including who they go to. Then build in slices that each go live and prove themselves, rather than one release at the end.
What you are left with.
A process that is faster because there is less of it, exceptions that route to a named person with the context attached, a record of what happened and who decided, and a design that can be changed by configuration when the policy changes.