A candidate asks who rejected their application. A data protection officer asks the same question, only more precisely. If AI is at work in your recruiting, the answer “a person decides” is enough only when you can show who decided and at which point.
The most common question about AI in recruiting is: who decides? The short answer is: a person. The trouble is that this sentence is easy to say and harder to show.
A lawyer and a data protection officer usually ask this question with Article 22 of the GDPR and human oversight under Article 14 of the AI Act in mind. This text is not legal advice. It shows how to map out the process so that the question “who decided” can be answered for every application.
A simple exercise helps. List all the points at which a candidate's application can end: after the form, after the rating of the application, after the call, after the task, after a deadline has passed. For each one, write down three things: who or what closes the application, on what basis and whether this can be reversed.
At each of these points one of three things happens. AI proposes. A person decides. Or the system closes the application on its own, following a rule set in advance. As long as these three things get mixed up in one sentence, you cannot tell who decided.
We will show this with the example we know best. The SmartyTalent virtual recruiter rates a candidate against the criteria from the list of requirements and prepares a recommendation. The criteria themselves are set by a person: the virtual recruiter writes down the manager's requirements, and you approve them before the job opens. Every rating refers to these criteria and has a reason.
After the rating of the application, after the call and after the task, the virtual recruiter proposes a decision and gives the reasons in plain language. It only proposes a rejection and gives the reasons. A person approves it: someone from your team or from our ReCops team, the people who supervise the virtual recruiter's work.
So that approval is not a formality, three things are needed in any tool. The person who approves sees the reasons, not only the result. They can disagree: change the rating or keep the candidate in the process. And they are a specific person who can be identified later. If any of these things is missing, approval becomes a click, and in practice it is the system that decides about the candidate.
In many processes there are also points where nobody approves anything. An honest description of the process names them plainly instead of hiding them behind a general sentence about oversight.
In our product only one situation closes an application without a person: an answer in the form that does not meet a must-have requirement agreed with the manager, for example holding a category B driving licence. Such knockout questions are stated openly: the candidate sees them before sending the application. A person can reverse a closure that was based on an answer in the form. When a candidate does not reply by the set deadline, for example to an invitation to a test, the virtual recruiter proposes closing the application and a person approves it.
For a rejection we record the requirement, the expected answer and the candidate's answer.
Such points are worth looking at most closely, because there is no person here who approves. For each one, ask who set the rule, whether the candidate knows it in advance, what is recorded and who can reverse the closure.
An approved rejection is not yet a message that has reached the candidate. Something is decided along the way too: the content, the moment of sending and whether someone reads the message before it goes out.
In our example every stage has a switch: automatic, manual or off. You set it. A rejection message goes out with a delay and during working hours, or waits as a draft for approval. In manual mode every message waits for approval.
You also set the wording of messages to candidates, their tone and how much they explain. The reasons given to a candidate refer only to the criteria of the position from the approved requirements, never to personal traits.
What the candidate knows about the involvement of AI belongs to the same picture. With us, the candidate knows from the start that they are talking to an AI and can ask for a person at any time. Before the call they hear an announcement that the call is recorded and that a virtual AI assistant is conducting it. If they do not agree to the recording or prefer to talk to a person, a person calls them. They choose the time of the call themselves, from three proposed in a text message.
The question about a decision usually comes back after some time: when a candidate asks for an explanation or when someone checks the process. That is when the record counts. In our example the history of every decision stays in the system: who, when and why.
Whatever the tool, check whether your record answers these three questions for every application that has ended. “Who” means a person's name and role, not the name of a department. “When” lets you establish what that person saw at the time. “Why” means reasons that can be shown to the candidate and the manager.
If rejections are approved by a team from outside your company, you have all the more reason to establish who they are and what rules they work by. The record should show these people too.
Film (in Polish): “Who decides? A person.”
AI can prepare a rating, the reasons and a draft of the message. The line between a proposal and a decision lies where a specific person approves the rejection: a person decides.