Is clicking "approve" enough to supervise?
Track 01 · Managers and Business Leads · English edition
What you will learn to do
design an approval that allows for a real examination of the proposed action.
Fictional pedagogical case: a premium services company is preparing to automate its refunds.
The course
The assistant proposes a refund of 180 euros. The screen displays a reassuring summary and an approval button. The manager must open several applications to find the original amount and the rule invoked.
Control exists within the workflow, but the interface makes exercising it costly. Human validation provides little value if the validator lacks the necessary elements or cannot modify the decision.
The approval screen must present the exact action. In our case, the manager must be able to identify the reservation, the beneficiary, and the amount. They must be able to find the rule used as well as any reported exceptions or uncertainties.
The assistant's explanation is not enough to prove that the reasoning is correct. The manager must be able to examine the decisive documents and data. A persuasive justification can be erroneous.
The level of control must be adapted to the action. A prepared response without sending and a money transfer do not have the same consequences. Explicit rules must determine when an approval is necessary and who can grant it.
Finally, the system must allow for refusal, correction, or a request for further review. Supervision must be able to change the outcome.
Your decision
The screen only displays "Eligible client, refund recommended: 180 euros". Propose the information to be added and a test for human review.
Formulate an answer before opening the answer key.
Open the reasoned answer key
Add the case identification, the beneficiary, the amount paid, the applicable rule, and its validity. Make the evidence accessible. Flag missing data and relevant exceptions.
In a training environment, present a proposal with a discrepancy between the displayed beneficiary and the one in the case. Observe whether the participant detects it and blocks the action. The test measures behavior in this scenario, not a general guarantee of vigilance.
Production to keep
approval information and the possibility of refusal or correction.
Evaluation criterion
The participant verifies the decisive data and can block the action.
Sources and references
- NIST, AI RMF 1.0, 2023
- OWASP, LLM06:2025 Excessive Agency
- Check the box! Automation bias in AI-based personnel selection, 2023
The case and the answer key are VOCE pedagogical constructions; the references do not validate the effectiveness of this training.