Where a person is needed, and where they are not
The right question is not “how much can I automate?”, but “where can I afford not to look?”
There are decisions a system can make autonomously: classifying a request, entering a standard order, answering a recurring question. And there are others where the cost of an error is too high to take the risk: a contractual exception, a sensitive complaint, an unusually large amount.
Distinguishing between the two is a risk management decision, not a technical one. That is why we make it together with you, process by process, before configuring any automation.
Two ways of working, not two philosophies
- Human-in-the-loop: the system prepares the response or action, and a person confirms it before execution. You save time while maintaining control.
- Human-out-of-the-loop: the system makes and executes decisions autonomously, within limits you define. You gain scale where the risk is low.
The level of autonomy is not an on/off switch, but a configurable parameter for each type of case: the same process can be fully automated for routine cases and fully supervised for exceptions.
How the system decides when to stop
Every automated decision comes with a confidence score. Below a defined threshold, the case is not executed but placed in an operator queue, with the proposed response and the reason behind it already prepared. Above the threshold, the case proceeds automatically.
Routing rules do not depend solely on technical confidence, but also on business criteria that you define: an amount above a certain threshold, a customer belonging to a specific category, a request containing certain topics, time of day, or workload. A case can be technically simple and still require human approval.
The adoption path
- Observation: the system makes a proposal, but no action is executed automatically. We only measure the level of agreement between the proposal and the human decision.
- Full supervision: every action requires confirmation, but the proposal is already prepared. The time savings begin here.
- Selective autonomy: categories with consistently high agreement are processed automatically, while the rest remain under supervision.
- Sample-based supervision: periodic checks are performed on a subset of cases rather than every single one.
None of these steps is irreversible: the threshold can always be rolled back.
Everything tracked, for audit and improvement
Every decision — automated or supervised — generates a record containing the input received, the proposed action, the confidence score, the rule applied, the outcome, and, in the case of human intervention, who intervened and what they changed.
This serves both audit and continuous improvement: operator corrections are the most reliable source for understanding where the system makes mistakes and where it can instead be trusted to operate autonomously.
A requirement, not just a best practice
The European Union’s Artificial Intelligence Act requires effective human oversight for systems classified as high-risk: the designated person must be able to understand the output, interpret its limitations, and intervene or stop the system.
An architecture that exposes reasoning, confidence scores, and correction capabilities is therefore not just a prudent design choice — it is a technical prerequisite for compliance. Risk classification must be assessed for each specific use case, and we evaluate it with you during the analysis phase.