Intelligent automation fails when teams try to make a messy process look smart. The more reliable sequence is to clarify the workflow, automate the predictable steps, and then assist the exceptions.
If the same case can take five different paths depending on who receives it, a model will inherit that inconsistency. Process design is still the first lever.
Once the happy path is stable, AI can help with documents, classification, and routing. Those are interpretation problems. They are different from moving a status field from one system to another.
Keep people in the loop where the cost of a mistake is high. An exception queue with context is often more valuable than a fully closed loop that nobody trusts.