The European AI Act (AI Act) classifies AI systems by risk. Part of the applications that relate to personnel falls into the high-risk category. That is not a qualification of how well or poorly a system works, but a classification based on what the system decides or advises, and about whom. Systems that concern the recruitment, promotion, evaluation, or dismissal of people touch on a fundamental right — the right to fair treatment in the employment relationship — and therefore receive a stricter regime: additional documentation, human oversight, and in many cases an obligation to inform the people involved.
This page describes the mechanism as it has been set up in the regulation. The complete, current text with the precise annexes and definitions is available on the website of the European Union and of the national government; that is where you will find the wording that carries legal weight.
The AI Act cites, as examples of high-risk applications in the employment context, systems that among other things:
The common denominator is that the system generates an outcome that has a direct effect on the position of an individual employee or applicant. A planning tool that optimises schedules for an entire department usually does not fall under this; a tool that generates a score per employee which is factored into a performance review does. Exactly where the line runs within your own processes is a question for a lawyer or compliance adviser, not something this page determines for you.
This is the part where many organisations get stuck, and it is also precisely the question that the three categories of this tool try to order. A task that AI can fully take over is generally a task without a lasting judgement about a person: summarising text, calculating a schedule, optimising a planning. As soon as the system produces output that serves as input for a decision about an individual — hiring, not hiring, promoting, not renewing a contract — the task shifts to the second or third category: human oversight that approves or rejects with reason, or work done entirely by humans.
The practical characteristic of a tool is that a human checks the input, can disregard the output without consequence, and makes the final judgement based on more than the system's outcome. The practical characteristic of an assessment system is that the outcome is given significant or decisive weight in a decision about that specific person, even if a human formally says "approved." A score that in practice is always adopted functions as an assessment system, regardless of what it is called internally. This delineation is comparable to what applies with the GDPR when automating tasks: there, too, it is not the label of the tool that counts, but what the tool actually decides and which data are used for that.
If a system qualifies as high-risk, this brings obligations for the party deploying it: a risk assessment, documentation on how it works, human oversight that can genuinely intervene, and transparency towards the employees or applicants concerned about the fact that an AI system plays a role. This page does not state percentages, deadlines, or amounts attached to these obligations, because these differ per situation and per phase of the regulation. Consult the official texts, or an adviser familiar with the current state of affairs, for the exact requirements and dates.
It is also important that this classification is separate from personnel decisions. The AI Act says something about the system, not about what an organisation does with the outcome of a reorganisation or role change. Decisions that affect the employment relationship are subject to their own legal requirements — think of dismissal law and the role of employee representation — which must be assessed separately and prior to the deployment of a system. This page makes no statement about that.
The introduction of a system that monitors or assesses the behaviour or performance of employees is often not only an AI Act question but also a question of employee representation. Consent rights can already apply before a system is classified as high-risk. When the works council must be involved in this and at what point is described on when the works council is involved. That process runs alongside, not instead of, the obligations under the AI Act.
This page describes the mechanism: which applications are mentioned, and where the practical distinction lies between a tool and an assessment system. Whether a specific system in your organisation falls under the high-risk category, and which steps follow from that, is a question that must be answered per situation — comparable to the nuances that also apply in what can AI take over in professional services, where the same question of oversight and responsibility returns.
Would you like a first impression of which part of the tasks in your job profile can be taken over by AI today, without looking at assessment systems concerning individuals? The free quickscan from ftetoai consists of twelve questions, works without an account, and gives an indication per profile. The full work scan, which goes deeper into individual processes and risk classification, is still under construction.
Vraag maar. Ik ken de kennisbank van deze site; wat ik niet weet, zeg ik erbij.
Answers come from this site’s knowledge base. Not tailored advice, and not a scan of your company.