Start with the decisions the work requires, then choose the simplest system that can make them reliably.
Map the inputs, decisions, tools, and exceptions before choosing a model. If the steps are predictable and the rules are stable, ordinary automation may be easier to test and maintain.
An agent becomes useful when a task requires interpreting context, selecting a tool, or deciding what to do next. Give it a bounded job, explicit permissions, and a clear route to a person when it cannot proceed.
Evaluate the whole workflow on representative examples. Look at completed tasks, incorrect actions, time, and cost together. A fluent answer is not enough if the next action is wrong.