Article 4 of the EU AI Act requires every employer and recruitment firm using AI agents to ensure the people working with them have sufficient AI literacy. In recruitment the bar is high: screening and ranking candidates is a high-risk use, so recruiters, hiring managers and agency staff must know what the agent decides and what they still own.
What does Article 4 ask of recruiters who work with agents?
Article 4 asks providers and deployers to take measures, to their best extent, so that staff and others dealing with AI systems on their behalf have a sufficient level of AI literacy. Sufficient is judged against the person's knowledge and experience, the context of use, and the people the system is used on.
In recruitment the people affected are candidates, often many at once, with no view of how they were assessed. That context pushes the bar up. A recruiter who approves a shortlist built by an agent needs to know how the shortlist was built, not just how to click approve.
Recruitment is also one of the uses the Act classes as high-risk when AI is used to filter applications, evaluate candidates or decide who moves forward. That brings in Article 14 human oversight and the deployer duties in Article 26, both of which assume the people overseeing the system are competent. Literacy is the floor those duties stand on.
Who in a hiring process does it cover?
More people than most HR teams expect. In a typical flow where agents help hire, the duty reaches:
- Recruiters and sourcers who brief the screening agent, set criteria and review its shortlists.
- Hiring managers who receive ranked candidates and decide who to interview.
- Recruitment coordinators whose scheduling agent books interviews and sends confirmations.
- HR operations who approve offers or rejections the agent has drafted.
- Agencies and outsourced recruitment providers screening on your behalf. Article 4 covers people dealing with the system for you, not only employees.
- Interview panels who read agent-written candidate summaries before they meet anyone.
The question of whether staff who supervise agents are in scope is settled: they are, and in recruitment they carry most of the risk.
What must recruitment staff understand about the agents they use?
Generic AI awareness does not help a recruiter at the moment they accept a ranked list. Literacy for recruitment agents is specific.
What the agent may do, by capability. A screening agent might read CVs at depth 1, draft a shortlist with reasons at depth 2, and book interviews at depth 3 within the calendar limits you set. It should never send a rejection on its own. Staff need to know which capability sits at which depth, so they notice when something arrives that should not have.
What the ranking rests on. An agent ranks on the sources it was given: the job description, the CV, the screening answers. A gap in employment, a non-standard degree title or a name the agent parses badly can move a candidate down for reasons that have nothing to do with the job. Recruiters must be able to read the agent's stated reasons and test them against the CV.
Proxies. Postcode, school, graduation year and career breaks can stand in for protected characteristics. Staff should be trained to spot criteria that act as proxies, in the brief and in the reasons the agent gives.
Automation bias. A neat, confident shortlist is hard to argue with. Recruiters need practice overturning agent recommendations on cases built to be wrong.
What held and refused mean. When the agent drafts a rejection and holds it for a named person, that person is making the decision. When the agent refuses to send an offer because it is outside its bounds, that is the system working, not a fault to route around.
Telling candidates. Candidates who chat with a screening agent must be told they are dealing with an AI system under Article 50. Coordinators need to know where that notice appears and what to do if a candidate asks for a person.
How do you meet the duty in a recruitment team?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Sufficient literacy for each role | Role map for recruiters, hiring managers, coordinators, agencies and panels, with a standard for each | Approved literacy policy with an owner |
| Account for the context of use | Training on the specific screening and scheduling agents in use, including their depths | Agent-specific module records per person |
| Account for the people affected | Content on candidate rights, proxies and fair treatment | Module content and assessment results |
| Oversight by competent people | Only trained, current staff may approve held rejections and shortlists | Approval log showing approver and their training status |
| Cover agencies and outsourced staff | Contract clause requiring role-based literacy and attestation | Supplier attestations and sampled work |
| Show it works | Seeded shortlists with a known error, reviewed by recruiters | Detection results by person |
| Keep it current | Refresher when an agent gains a capability or its brief changes | Change-linked training records |
Two rows matter most. The approval log ties each decision to a named, trained person, which is what a candidate complaint will test. The seeded cases show the training changed behaviour, which completion records cannot. For the wider build, see how to set up an AI literacy programme under the EU AI Act.
What will a supervisor or auditor ask?
Expect questions like these:
- Which of your recruitment steps use agents, and who oversees each?
- What does the screening agent do without a person? Ask the recruiter, not the vendor.
- Show us the training record of the person who approved this rejection.
- How do you check that recruiters are not accepting shortlists unread?
- What do agency recruiters know about your agents, and how do you know?
- When the agent's brief changed, who was retrained?
Boards are now asked the same questions in summary. The argument that the EU AI Act made AI literacy a board question applies with force where hiring is involved.
What does Article 4 not cover in recruitment?
It does not prescribe a course, hours or a certificate. It does not on its own make your screening lawful: the high-risk duties, data protection law, including Article 22 GDPR on solely automated decisions, and employment and equality law all apply separately. It does not transfer responsibility to a vendor; a deployer must decide what its own staff need. And it does not stop at the people who build agents. The coordinator who forwards an agent-drafted summary to a panel is in scope.
Other sectors face the same duty with different weight. Compare how it reads for AI agents in the public sector, where hiring rules meet public law, and for AI agents in telecoms, where the affected people are subscribers rather than candidates. Our guide to AI literacy and certification for people who work with agents sets out the full picture.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where the Academy fits
Pryme Intelligence Academy trains recruiters and HR reviewers on the real product. Foundations covers the decision chain and depth for anyone in the hiring process. The Operator track teaches human review and when something goes wrong, with module tests set in the learner's own industry and role. Certified Pryme Operator requires a written exam and a two-week practicum of real review decisions signed off by a Pryme assessor, and each certificate can be checked at source through VerifyDoc.ai.
Questions people ask
Does Article 4 apply to a small employer that uses one screening agent?
Yes. It applies to every deployer regardless of size. A small employer can meet it with a short, role-specific standard and records, but it cannot skip it.
Can we rely on the agency's training for agency recruiters?
You can use it, but you must check it covers your agents and your process. Put the requirement in the contract and sample the agency's shortlists as you would your own.
Should hiring managers who never touch the agent be trained?
If they receive agent-ranked candidates, yes. They act on the output, so they need to know what the ranking rests on and when to ask for the full list.