Screening applications under the AI Act: what an AI agent may do

An AI agent may prepare: read applications, compare them with the job description, argue a shortlist and write comparable interview summaries. It must not decide alone: recruitment is a high-risk use under Annex III of the AI Act, with human oversight, information for candidates and retained logs, applicable from 2 December 2027 according to the digital omnibus, a date to check when you read this.

This is not legal advice. The dates and articles quoted were checked on 30 October 2026; the texts are listed at the end. Check them again when you read this, and have your situation reviewed by your lawyer or data protection officer.

Why recruitment is high-risk

Annex III of the AI Act lists the areas where an AI system is presumed high-risk. Point 4 is titled “Employment, workers’ management and access to self-employment”. It covers, first, systems intended for the recruitment or selection of people, in particular to target job ads, analyse and filter applications, and evaluate candidates. It covers, second, those used to decide on a promotion or a termination, to allocate tasks based on behaviour, or to monitor performance.

The logic is easy to follow. A candidate set aside at first reading will often never know why, and her career may depend on it. The regulation therefore places these systems under strict requirements, and it gives much of the work to the employer that uses them.

The timeline has moved. According to the European Commission page opened on 30 October 2026, the rules for high-risk systems in the sensitive areas of Annex III, employment included, will apply from 2 December 2027, and those for systems embedded in products from 2 August 2028. Our article on the AI Act and agents uses the same date. It does not excuse you from starting: the AI literacy obligation of Article 4 and the prohibited practices already apply.

Provider or deployer

A company that uses a purchased agent under its own authority, for its own hiring, is as a rule a deployer. That is the role Article 26 addresses, with the obligations described below. The provider answers for the design of the system and its documentation.

One detail changes the picture. If you substantially modify the tool, or change its purpose so that it becomes high-risk, you may move to the provider side (Article 25). A writing assistant that a team repurposes one day to rank candidates is an example: the tool was not meant for that, and you made that call.

What the agent prepares

A Tuesday, 8:40 a.m. In a mid-sized logistics company, the head of recruitment opens her inbox: 214 applications for three transport coordinator positions, received over the weekend. She has an hour before her first meeting. The agent read the files overnight. It presents each application in a few lines, links each criterion of the job description to what it found in the CV, or says it found nothing, and proposes a shortlist of twenty files with the reasons.

That is the work an agent does well: reading fast, comparing with a grid the team wrote, drafting. In the same vein, it formats interview summaries structured the same way for every candidate, which makes comparison possible, and it prepares replies to candidates, acknowledgements or reasoned refusals, that the recruiter reads before sending. That is the scope Scalt’s HR page describes: an argued shortlist against the job description, the decision stays human, and comparable interview write-ups.

The context explains the appeal. According to France Travail’s labour needs survey, run from October to December 2025 across 416,588 establishments, 2.28 million positions are to be filled in 2026, and 43.8% of recruitment projects are judged difficult, against 50.1% in 2025. Pressure on recruiters’ time remains high, and that is exactly where the temptation to hand over the decision appears.

What it does not do

The agent prepares, it does not decide. In practice, it sets no candidate aside on its own initiative: every refusal goes through a person who has seen the file. It sends no final reply without approval. And it does not rank candidates on criteria nobody can read, such as an overall score with no explanation, because the oversight of Article 14 assumes you can interpret what the system produces.

That article provides that the people in charge of oversight can understand the system’s capacities and limits, stay aware of the risk of relying too much on its output, interpret what it produces correctly, and decide in any given case not to use it, to disregard it or to reverse it. An argued shortlist gives the recruiter the means to exercise that power. A rank with no reason deprives them of any way to do so.

A word on the boundaries. The regulation provides that an Annex III system is not high-risk if it poses no significant risk, for instance because it performs a task preparatory to a human assessment, but that exception does not apply when the system profiles people. An agent that ranks candidates looks very much like an agent that profiles them. Where exactly the line falls for your setup is a question for your lawyer, with a precise description of what the agent does.

The deployer’s obligations in practice

For a high-risk use, Article 26 asks the employer to use the system according to its instructions for use, to assign oversight to people who have the necessary competence, training, authority and support, and to monitor operation. If it sees a risk, it informs the provider without undue delay and suspends use. Here is what that looks like for a hiring process:

• Real oversight: the person who approves the shortlist has time to read the reasons, the right to amend the list, and the competence to judge the job description. Clicking “approve” on twenty files without opening them oversees nothing.
• Informing employees: Article 26(7) asks the employer to inform workers’ representatives and the affected workers before using the system at the workplace. National rules on information and consultation come on top.
• Informing candidates: paragraph 11 asks to inform the natural persons who are subject to a decision made or assisted by the system. In a job posting, one paragraph is enough, as long as it is clear.
• Log retention: paragraph 6 requires keeping the automatically generated logs, to the extent you control them, for at least six months, unless Union or national law provides otherwise.

Article 86 adds the right, for a person affected by a decision based on the output of an Annex III system, to obtain clear and meaningful explanations of the system’s role in the decision, where the decision has legal effects or significantly affects them and they consider it adverse to their rights. A rejected application fits. If the agent left a trace of its reasons, the answer already exists; if not, it has to be reconstructed. Article 27, the fundamental rights impact assessment, targets mainly public bodies and public services: in principle it does not concern a private employer that recruits.

When a company gets equipped, the question quickly boils down to this: who on the team can show, afterwards, what the agent read and proposed? In Scalt, every shortlist opens an execution case (Cases) that keeps the request, the sources read, the reasons proposed and the recruiter’s approval, with their name and the time. That provides the material for logs and explanations. Retention periods and informing people remain the company’s decisions.

Candidates’ data

A CV is a document full of personal data, sometimes sensitive data when it mentions a disability or union involvement. The GDPR applies at the same time as the AI Act, and its Article 22 frames decisions based solely on automated processing. Your data protection officer will want three answers: how long you keep the files of unsuccessful candidates, how a person exercises their right of access, and what the model actually sees.

On the last point, personal data is anonymised before any model call in Scalt, and the agent reads only the sources the team has opened to it, with the rights of the person it works for. A recruiter who cannot see another department’s files does not give their agent access to them. As for the retention period, set it, write down why, and have it validated: no tool chooses it for you.

The register and the kit

Entering the screening agent in a register of AI agents is the simplest way to turn these obligations into habits. Scalt’s AI Act kit contains an Excel register that suggests a risk level for each agent and dates its reviews, along with a four-question classification grid and the deployer checklists. It helps organise the work; it does not do it for you and is not legal advice.

Whether you fill in the kit’s register or your own spreadsheet, the screening agent’s line should say in one sentence what it does, for which role and which teams, which sources it reads, who oversees it and with what authority, where the logs are and for how long, how candidates and employees are informed, and the date of the next review. If one of these boxes stays empty, you have just found the work that comes before December 2027.

Frequently asked questions