What can it decide?
That is the first question.

Spot the decisions, not just the steps

Write down the current flow with its branches. A stable condition, such as the presence of a reference, can often be checked by a rule. A workflow fits when the steps are known but some of them call for interpretation. An agent becomes interesting when it has to choose which information to look for or which tool to use depending on what it finds.

Flexibility has a cost: you need to be able to explain the paths taken and check more situations. Do not ask for it on a step that mainly benefits from consistency.

Break down a concrete delegation

Fictional example: an assistant prepares the comparison of three supplier proposals. It can look up the authorized documents and fill in a grid. If an annex is missing, it flags the absence. It does not contact the supplier on its own and does not turn a preference in the text into an approved order.

For each operation, define what can be executed directly, what remains a draft, and what requires an identified person. A useful approval shows the differences and the uncertain information; an “approve” button in front of an unreadable file protects very little.

Test where control must come back

Replay a complete case, a contradiction between two documents and an unavailable tool. Observe whether the system stops at the right place, keeps what it has already prepared and asks for specific help. Also check that a restart does not repeat an action already carried out.

Compare the workflow and the agent on the same cases. If the second multiplies searches without improving the accepted result, the extra freedom does not yet serve this flow.

Widen one operation at a time

Start with the level that allows real use with an understandable recovery. After observation, a repetitive, well-controlled operation can be delegated further, while an exception stays manual. Autonomy thus becomes a working setting, not a general promise.