Skip to content

The AI Act in recruiting: what HR should know before deployment

Questions for an AI recruiting vendor and things to prepare at your company before deployment, for HR, lawyers and data protection officers.

It sometimes goes like this: HR chooses an AI tool for recruiting, and the lawyer and the data protection officer get the contract at the end, when the decision has already been made. At that point it is hard to ask about anything, and even harder to change anything. So I have gathered what is worth asking the vendor and what to prepare at your company before the system starts working.

One classification and three dates

First, a caveat: this text is not legal advice. It is a list of questions and things to prepare. Leave the assessment of your own situation to your lawyers.

Under the AI Act, that is, the EU's Artificial Intelligence Act, systems for selecting and rating candidates are, as a rule, high-risk systems (Annex III, point 4, and Article 6). That is why I start with the question of how the vendor classifies its system and which documents you will receive.

It is also worth knowing three dates. Emotion recognition in the workplace has, as a rule, been prohibited since February 2025 (Article 5 of the AI Act). From August 2026 a person must be told that they are talking to an AI, at the latest at the start of the conversation (Article 50 of the AI Act). The requirements for high-risk systems in recruiting apply from 2 December 2027 (Regulation (EU) 2026/1744).

Two questions for the vendor follow from the first two dates: whether the system recognises emotions or analyses voice and facial expressions, and how the candidate finds out that they are talking to an AI. The third date looks distant, but you are signing the contract today. So it is worth asking about the documents and the division of roles now.

The provider and the employer are responsible for different things

I often hear the question of who is responsible for the obligations under the AI Act: the provider or the employer. The answer has two parts.

The provider of a high-risk system is responsible, among other things, for risk management, data quality, documentation, the instructions for use and the conformity assessment. The employer that uses this system follows the instructions, ensures human oversight, keeps the logs and informs workers, their representatives and candidates.

Three points on the employer's side are worth reading more closely. Oversight is carried out by people with competence, training, authority and support (Article 26(2) of the AI Act). The employer keeps the logs for at least 6 months, if they are under its control and other laws do not provide otherwise (Article 26(6) of the AI Act). The employer informs workers and their representatives before the system is used, and candidates when the system helps to make decisions about them (Article 26(7) and (11) of the AI Act).

There is also a trap. The employer may itself become a provider, for example when it makes a substantial modification to the system or puts its own name or trademark on it (Article 25(1) of the AI Act). Before you ask the vendor for a version under your own name, ask a lawyer what that changes in the division of roles.

So there are three things to prepare at your company. Name the people who will oversee the system and plan their training. Agree with the vendor who keeps the logs and for how long. Plan when and how you will inform workers, their representatives and candidates. It is worth writing the division of roles into the contract.

A person decides, and the candidate knows who they are talking to

The next topic is decisions about candidates. Article 22 of the GDPR concerns decisions based solely on automated processing. So ask the vendor to show what the system does on its own, and where a person really assesses the result and does not just formally approve it.

I will show this with the example I know best. In the SmartyTalent virtual recruiter a person sets the criteria: the brief from the conversation with the manager has to be accepted before the work starts. Every rating of a candidate refers to these criteria and has a reason. A rejection after the rating of the CV, after the call and after the task is a proposal with a reason, and a person approves it. Only one situation closes an application without a person: an answer to a knockout question stated openly in the form, for example about a category B driving licence, that does not meet a must-have requirement agreed with the manager. We treat this as an automated decision, and a person can reverse such a decision. 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. The history of every decision stays in the system.

The other side of the same matter is the candidate. When an AI calls them, I check three things. First, whether the candidate knows that they are talking to an AI. Second, whether they know that the call is recorded, what for, and how long the recording will be kept. Third, whether they can refuse and talk to a person. In our example, before the call the candidate hears that the call is recorded and that an AI is conducting it. If they do not agree to the recording or would rather talk to a person, a person calls them.

Documents and the contract before the decision

It is worth getting the answers in writing, together with the documents. Before you choose a tool, ask the vendor for:

  • a description of how the system operates: how it works and what its limitations are,
  • the instructions for use: how to use the system and how to oversee it,
  • a risk assessment with a discrimination analysis, and a description of the data the system was tested on,
  • materials for your data protection impact assessment (DPIA),
  • a data processing agreement with a list of sub-processors,
  • a training programme for the people who oversee the system.

The data protection impact assessment is carried out by the employer as the data controller, and the vendor supplies information (Article 35 of the GDPR).

Then the contract. The answers from the presentation must go into it. Otherwise they will remain a promise. In the part about data I check three clauses. The first is the data processing agreement, with information about every change of sub-processors and the right to object. The second is the countries in which the data is processed, and the basis for transfers outside the European Economic Area. The third says whether candidates' data may be used to train models. The next three concern security and the end of the cooperation: the time limit within which the vendor informs you of a breach, the right to audit with access to the system's documentation, and the return or deletion of the data after the contract ends.

There are also three clauses related to the AI Act. The first is the instructions for use, with the information a person needs to oversee the system. The second is the logs: who keeps them and for how long. The third is changes to the system: how and when the vendor informs you of those that affect the rating of candidates.

At SmartyTalent we hand over the set of documents before the decision on deployment. It includes, among other things, the technical documentation, model cards and data cards, a risk assessment with a discrimination analysis, the instructions for use and a training programme for the people who oversee the system. We already document the system today with the requirements for high-risk systems in mind.

To check at your company

  • How does the vendor classify its system under the AI Act, and which documents will we receive before signing the contract?
  • Who really makes the decision about a candidate, and what does the system do on its own?
  • Who at our company will oversee the system, and do they have the training, authority and support for it?
  • Who keeps the system's logs, and for how long?
  • How does the candidate know that they are talking to an AI, and how can they ask to talk to a person?

For further reading I recommend the guide from our blog (in Polish): “Free guide to the AI Act and GDPR for HR”.

Film (in Polish): “Security and the AI Act in the virtual recruiter”

AI can prepare a recruitment process, but it is worth having the questions, the documents and the contract ready before the AI starts work. Then the tool prepares, and a person decides about the candidate.

Book a 30-minute demo

Monika Wyszynska

Monika Wyszynska

Monika is a Chief Operation Officer at SmartyTalent, a technology startup that helps leverage the power of AI to Enhance Sales&Support meetings.

Latest Articles

ReCops: who supervises the work of AI in recruiting

ReCops: who supervises the work of AI in recruiting

Who watches AI when it starts working in recruiting? The tasks of supervision, two models for running it and the questions worth asking a v...

The AI Act in recruiting: what HR should know before deployment

The AI Act in recruiting: what HR should know before deployment

Questions for an AI recruiting vendor and things to prepare at your company before deployment, for HR, lawyers and data protection officers...

The candidate is waiting for an answer. What silence in recruiting costs and how to shorten it

The candidate is waiting for an answer. What silence in recruiting costs and how to shorten it

No reply after an application costs the employer brand. Where silence in recruiting comes from and how to shorten it when the team is short...