An agent can issue small refunds and also close accounts. How should controls differ?
Why B is the answer
Automate small reversible refunds, and require approval for closures
Match the control to the impact. Reversible, low-value actions can run with logging, and irreversible ones need a person.
In detail
- Match the control to the impact.
- Small, reversible actions can run with logging.
- Irreversible actions need a person.
- Set the limit with the business, and review it.
Use case
An illustrative scenario, not a real incident.
An agent refunds small amounts by itself, with logs and a daily sample. Closing an account always goes to a person, because it cannot be undone.
See it in different ways
flowchart TD
A{"Can the action be reversed?"} -->|Yes| B{"Low impact?"}
A -->|No| C["Human approval required"]
B -->|Yes| D["Automatic, with logging"]
B -->|No| E["Automatic, with review sampling"]stateDiagram-v2 [*] --> AssessAction AssessAction --> Automate: low value, reversible AssessAction --> HumanApproval: high risk, irreversible Automate --> Logged HumanApproval --> Logged Logged --> [*]
flowchart TD
Q{"What is the action like?"}
Q -->|"Small and reversible"| A0["Automate with logging"]
Q -->|"Irreversible"| A1["Human approval"]Key point: Match the control to the impact.
Why the other options are wrong
- A. Require a human to approve every action, including the small refunds
- Approving every small refund wastes reviewers' time.
- C. Let the model estimate its own confidence and act only above a threshold
- Self-reported confidence is not a reliable control.
- D. Apply the same automatic approval to both, so that the design stays simple
- A closure cannot be undone, so it needs more than the same approval.