Human Competence Gate – when AI checks whether you really understand what you are approving. Problem: "OK" without understanding

AI systems are increasingly supporting people in their decision-making – they analyze documents, assess risks, and recommend actions. Humans act as supervisors: they see the recommendation and click "Approve." That's the theory.

In practice, there is a serious gap. AI recommendations are based on hundreds of variables, dozens of documents, and complex legal or financial relationships. Humans see the end result – the proposed decision – but do not always understand why AI proposed it and what the consequences of approval are.

This creates three specific risks:

• The illusion of oversight – the user is formally "in the loop," but in reality is approving something they are unable to assess. The OK button becomes a reflex, not a decision.

• Audit gap – in the event of an error, it is impossible to prove whether the user actually understood the issue. The event log only shows that someone clicked "Approve" at 2:32 p.m.

• No verification mechanism – existing control systems (RBAC, roles, permissions) check whether the user has the right to approve a decision, but do not check whether they understand what they are approving.

The AI Act requires effective human oversight of high-risk systems. "Human-in-the-loop" alone does not guarantee this if the human in the loop does not understand what they are participating in.

Method: Human Competence Gate (HCG)

Human Competence Gate is a mechanism in which the AI system, before unlocking the user's ability to approve a decision, asks them a series of contextual "Yes/No" questions to check their understanding of the key facts, consequences, and limitations related to the specific case.

How it works

The AI generates a recommendation, but the approval button is blocked. Instead, the user sees several questions generated dynamically based on the content of the case – documents, legal basis, business rules, and decision context.

The questions cover three areas:

• Facts – e.g., "Did you know that the applicant has exceeded the acceptable debt ratio?"

• Consequences – e.g., "Does approving this decision involve a financial commitment for 5 years?"

• Restrictions and exceptions – e.g., "Can this decision be appealed within 14 days?"

Validation and result

The system compares the user's answers with the actual state of affairs. If the answers are correct, the OK button is unlocked. If the user answers key questions incorrectly, the system blocks approval and may suggest one of three actions:

• reviewing the case documents again,

• forwarding the decision to a person with higher authority,

• refer the user to the training module.

Audit trail

Each HCG session is fully logged: the content of the questions, the user's answers, the validation result, the time, the model version, and the final decision. This means that in the event of an audit, it is possible to reconstruct not only what the user approved, but also whether they understood what they were approving.

Potential: AI that improves humans

HCG solves the problem of supervision, but its true value lies elsewhere.

No more "rubber-stamping" decisions

Without AI support, people also make decisions based on intuition, habit, or a cursory reading of documents. No one checks them, no one asks, "Are you sure you understand?" HCG turns every decision into a micro-exam – it forces you to reflect on the consequences before your finger touches the button.

Education built into the process

HCG questions are not only for verification – they also serve an educational function. Users who regularly go through the competency gate learn facts, regulations, and relationships that they previously ignored. The system collects data on incorrect answers and can direct users to targeted training – precisely in those areas where their knowledge is weakest.

Measurable increase in competence

Traditional training courses end with a certificate and then are forgotten. HCG generates a continuous, measurable competency profile for each user – how many decisions they have approved, what percentage of correct answers they have given, in which areas they make mistakes, and how their effectiveness changes over time. This is data that no training system today provides.

Response to the accusation that "AI takes away competencies"

One of the most common objections to AI systems is that they take away people's skills – we delegate thinking and lose agency. HCG reverses this argument. Instead of replacing humans, AI becomes a mechanism that enforces and builds competencies. Paradoxically, the more AI is used in the process, the more competent the person using it becomes.

A new standard of supervision

HCG is not just about securing the process. It is a new standard for what "human-in-the-loop" should mean – not the presence of humans in the loop, but the proven competence of humans in the loop.

The Human Competence Gate (HCG) model is ready to be implemented as part of the AI Governance architecture in decision support systems in banking, public administration, HR, and medicine.