Start with one real request
Choose a recent piece of work and follow it from beginning to end. Write down who received it, which tools they opened, what they copied and where they had to wait. A concrete example reveals details that a generic process diagram often misses.
Separate work from waiting
A task that takes ten minutes may spend three days waiting for a decision. Automating the ten minutes will not fix the delay. Record both handling time and waiting time so that the next investment targets the actual constraint.
Look for a stable rule
Good early candidates have a recognisable trigger, a clear decision rule and an observable outcome. Copying approved invoice details between systems is different from deciding whether an unusual invoice should be approved. Keep uncertain decisions visible to a person.
Design the exception first
Ask what happens when a source system is unavailable, a field is missing or the same request arrives twice. Name the person who owns the exception and give them a useful recovery path. An automation that fails silently simply creates a new kind of manual work.
Run a small, measurable pilot
Choose a narrow workflow and agree on the evidence that would make it worth continuing. Measure a baseline, run with a small group and compare the outcome. Include maintenance time and the cost of exceptions in the assessment.