Catching Risk Before It Becomes an Incident

By the time AI-assisted work becomes an incident, the damage is already partly done. A harmful recommendation, a fabricated source, a piece of sensitive information handled the wrong way: once any of these reaches a client, the firm is into damage control. The more useful question is what happens before that point. Can the system notice that an answer is risky while there is still time to stop it, rather than only after it has gone out?

That is what risk screening is for. It is not a replacement for human review, and it is not a guarantee that nothing slips through. It is an early-warning layer that looks at AI output for the signals that tend to precede a problem, so the riskier answers get flagged for attention instead of flowing through at the same speed as everything else. Prevention is cheaper than incident handling, and screening is how prevention scales.

Why human review alone is not enough at volume

Human review is the heart of responsible AI use, but humans reviewing at volume have a predictable weakness: everything starts to look the same. After the fortieth routine answer, the reviewer is primed to approve, and the one answer that actually carries a risk does not announce itself. It looks like the thirty-nine that were fine. Attention is a finite resource, and asking a person to spread it evenly across a high volume of work is asking them to miss the rare dangerous case.

Screening helps by being tireless where humans tire. It applies the same checks to the fortieth answer as the first, without fatigue, and raises a flag when something matches a risk pattern. That does not make the decision for the human. It directs the human’s limited attention to the answer most likely to need it, which is exactly where review tends to fail without help.

Two tiers: cheap first, careful second

Screening everything with a heavy process would be slow and expensive, so Qonera screens in two tiers. A fast, zero-cost first pass checks every answer for risk signals: things like personal or sensitive information, harmful advice, fabricated sources, or security-sensitive content. Most answers clear this pass and move on with no added cost or delay.

When the first pass detects a signal, a more careful classifier takes a closer look and returns a structured judgment of the risk and its severity. The cheap check runs on everything; the expensive check runs only where the cheap check found something worth a second look. That keeps screening practical at volume while still applying real scrutiny where it is warranted, rather than forcing a choice between screening nothing and screening everything slowly.

From flag to action

A flag is only useful if something happens with it. When the screening detects a high-risk answer, it can feed into the approval gate so the answer stops for a human, and it can create an incident record so the event is tracked rather than lost. The screening is the early-warning layer; the review and incident workflow is what turns a warning into a decision and a record. See how the pieces fit on the workflow page.

The same principle sits behind Article 9 of the EU AI Act, which expects organizations using high-risk AI to identify foreseeable risks, evaluate them, and put controls in place to keep them within acceptable limits. Screening output for risk signals before it reaches a client is a practical version of exactly that. Most of the obligations under the EU AI Act apply from August 2026, and a team that already catches risky answers before they become incidents is working the way the risk-management expectation points. The goal is not to eliminate human judgment. It is to make sure the human is looking at the right answer at the right moment, before a risk becomes a problem the firm has to explain after the fact.

This article is for general information only and does not provide legal advice. Organisations should consult qualified legal counsel about how Article 9 and the EU AI Act apply to their specific systems, workflows, and obligations.