Human validation for sensitive decisions and full control over the actions of AI agents.
The Inbox is each user’s task list in VOGO eBS AIAO: the workflow steps that need a person arrive here — an approval, filling in a form, an assignment, a reply.
What can be done from a task
- Assignment to a department or to a person in a department; if the step has a defined scope, only the structures and people in that scope can be chosen.
- Discussions on the task, with the messages of those involved, kept next to the task.
- Four-eyes approval: the task shows how many approvals are needed and who has already approved, with the date.
- The task’s deadline and its history, from creation to closing.
- The documents of the case, consulted from the task, within the user’s permissions.
Automation without brakes is a liability, not an advantage
A system that decides alone on matters affecting people moves responsibility to a place where it cannot be exercised. Who answers for a wrong classification that sent a petition to the wrong department and missed the thirty-day deadline?
Regulation (EU) 2024/1689 on artificial intelligence formalises this intuition: for high-risk systems, human oversight is not a good practice but a requirement. And in public administration, many workflows fall into this category.
How people stay in the loop
The level of autonomy is set per type of operation, not globally. Proposals without direct consequences can be applied automatically; those that produce effects — a registration, a classification with a legal deadline, a rejection — can require confirmation.
When confirmation is needed, the case reaches the right person with the full context: the document, the agent’s proposal and its reason. The person confirms, corrects or rejects, and their decision stays on record next to the original proposal.
Corrections are not lost. Over time, they show where the agent errs systematically and where the rules need adjusting.
- autonomy configured per type of operation, not globally;
- the case reaches a person with the proposal and its reason;
- the human decision stays on record next to the agent’s proposal;
- corrections show where the rules need adjusting.
What the organisation gains
Automation you can defend. At an inspection, an appeal or an internal discussion, you can show who decided, on what basis, and that the decision was not left to an unsupervised system.
And a practical, less obvious effect: the team’s trust. A system that asks for confirmation where it matters is accepted more easily than one that decides alone and has to be corrected afterwards.
Questions and answers
Who decides which decisions need confirmation?
The organisation, through configuration. The level of autonomy is defined per type of operation.
Doesn’t this slow automation down?
Confirmation is asked only where it has been configured. The rest of the workflow stays automatic, and the case reaches a person prepared, not from scratch.
What happens with the decisions people correct?
They stay on record, next to the original proposal, and are used to adjust the rules.
How is human oversight shown at an audit?
Through the audit log, which holds both the agent’s proposal and the decision of the person who confirmed or corrected it.