An automated system can prepare a useful result and still take the wrong action with it. Drafting a customer reply is different from sending it. Identifying an outdated file is different from deleting it.
The useful habit: place review before the consequential action, when the reviewer can still change the outcome.
Map preparation, decision and action
Imagine a fictional support workflow: read a request, summarise it, propose a refund, update the account and notify the customer. These steps do not have the same consequences. A poor summary can be corrected internally; an incorrect refund or promise can affect money and trust.
Write down where information is merely prepared and where the outside world changes. Those boundaries are natural places to consider approval, validation or a restriction on what the system can do.
Choose review points by consequences
| Workflow step | Possible control |
|---|---|
| Format an internal draft | Automate within defined constraints and sample-check |
| Interpret an ambiguous request | Ask for clarification or route to a reviewer |
| Make a commitment to a customer | Show the exact commitment before sending |
| Change access or sensitive records | Require an authorised reviewer and enforce permissions |
| Perform an irreversible operation | Use a specific approval, recovery planning and appropriate technical safeguards |
This table is a design example, not a universal approval policy. Your organisation’s requirements, the system’s demonstrated reliability and the consequences of error determine the actual controls.
Make the approval concrete
“Continue?” is a poor review request if the person cannot see what will happen. Show the proposed action, recipient or destination, relevant data, reason and important uncertainty.
For the fictional refund workflow, the reviewer should see the order, amount, policy basis and exact customer message. Approval of one amount should not silently authorise a different amount after the underlying data changes.
Bind approval to the specific action and parameters. If the proposed action changes materially, the previous review may no longer apply.
Give the reviewer a real choice
A useful checkpoint supports approve, edit, reject and request more information. The reviewer needs enough time, knowledge and authority to choose. A person clicking through hundreds of vague alerts may provide little meaningful oversight.
Reduce unnecessary interruptions by automating well-defined, low-consequence preparation. Escalate uncertainty and exceptions clearly. This helps reserve attention for the decisions where it matters.
Approval is only one layer
OWASP’s guidance on excessive agency emphasises restricting available functions, permissions and autonomy, with human approval for high-impact actions. A model should not receive broad account access merely because a person may review its work later.
Enforce limits in the systems that execute actions. A read-only task should not require deletion privileges. A workflow proposing an action should not be able to bypass its own approval gate by using another tool.
Design the failure path
- Define stop conditions. Missing data, conflicting instructions and unexpected tool results should have explicit handling.
- Record what happened. Keep appropriate logs of the request, proposed action, approval and actual result without unnecessarily exposing private information.
- Verify execution. A submitted request is not proof that an update succeeded.
- Plan recovery. Identify who can pause the workflow, reverse a reversible change or handle an incident.
- Test the checkpoint. Include changed parameters, expired approvals and rejected actions in the test cases.
Review the controls as the workflow grows
A process that once drafted internal notes may later gain email access or payment capabilities. That changes its risk even if the original prompt stays the same.
Revisit approval points when tools, permissions, audiences or data types change. Use a small test set to expose ordinary failures before expanding access.
Note to keep: good human oversight gives a capable person clear evidence and a real opportunity to intervene before the consequence occurs.
Related: compare AI answers and protect information before an AI upload.
