An agent can suggest a change to a supplier's data, or it can have permission to save that change in the ERP. In the second case, checking whether the suggestion was right is only part of the job. Someone has to decide who authorizes the change, what gets logged, and how to correct an error or undo the change.
In practice those decisions tend to be split between the business team, IT and the implementation partner. Responsibilities should be clear across those teams before an agent gets write access, not after.
A reasonable starting point is a list. For each AI tool in use in EBS or Fusion, write down what it can do (query data, propose a change, execute it), who approved that permission, and who gets called when it fails. In Fusion, the review should include the AI features enabled in the environment. In EBS, it's also worth covering custom integrations and any agent connected to the system.
The review should also cover how people approve the agent's suggestions. Does the approver have enough information to spot a wrong supplier or an amount that doesn't add up? Can they tell which suggestions were generated by AI? To review a change, the approver needs to see which data will be modified and what information the proposal was based on. A suggestion and an approve button may not be enough to catch an error.
With that inventory in place, you can check whether each agent's permissions match its task and define who can extend them. It should also record who to contact and how to stop execution if an error shows up. In our agent deployments on Oracle, we document these points before enabling permissions to modify data.