Imagine that in a Libreville bank, an artificial intelligence assistant prepares a vendor transfer, checks the authorized ceiling, then requests the payment to be triggered. The file appears clean. Yet, one question blocks everything: who actually authorized the action, and what proof will the audit be able to read in three months?
In a university, the same problem arises with scholarships. A tool sorts applications, identifies missing documents, suggests a priority, and prepares notifications. If the rule applied is wrong, or if personal data is used without a clear basis, the speed becomes a compliance risk.
What is it, concretely?
An artificial intelligence agent (AI agent) is not just a program that answers a question: it can prepare an action, such as classifying a file, producing a report, issuing an alert, or triggering an operation. The Safeguards for Agentic Finance at Runtime (SAFR) framework, published with the Monetary Authority of Singapore, proposes placing a control between the AI agent and the business system. Before the action, the organization verifies who is acting, what is authorized, which rule applies, if a human needs to validate, and what record will be kept. The key point is simple: AI can speed up work, but responsible decision-making remains organized, controlled, and auditable.
Concrete case: what to do and what not to do
Questions to Ask Before Acting
Does the AI offer a recommendation or can it act directly in the system?
Who bears responsibility if the action is false, unjust, or non-compliant?
What actions should remain forbidden for AI, even if they seem useful?
From what risk level is formal human validation required?
Is it possible to find out why a decision was made three months later?
Are the data used necessary, authorized, and protected?
Can internal audit test the system with concrete evidence?
UNIVGA Press
French 
English