Automation and agent are not the same thing
A simple automation follows a path we already know: it receives a piece of data, applies a rule and returns a result. An agent receives an objective, chooses the steps and uses tools to reach it. The practical difference is who decides: in the first case the process we drew decides; in the second, the model decides at the moment of execution.
When an agent makes sense
It makes sense when the path is not fixed but the objective is. Classifying emails and answering the ones that can be answered, searching several systems to assemble a report, or reviewing a request and routing it with a summary. In all three cases the person can describe what they want to obtain and does not mind that the order of the steps changes from case to case.
The risks to bound beforehand
An agent that decides needs explicit limits. We always define three: which tools it may use and which it may not, which actions require human approval before running and how far it may spend. On top of that every step it takes has to be logged, because an agent that cannot be audited cannot be corrected either.
How to start without breaking the operation
We start with the agent in read-only mode: it observes, proposes and does not execute. For a few weeks we compare its proposals with what the human team does. When the rate of agreement is high we let it execute the reversible actions and keep human approval on the ones that are not. That way the team builds trust from evidence and not from slides.
