Share
How the EU AI Act changes recruitment in Romania: what employers can and cannot do
You have an ATS (applicant tracking system) that filters CVs by keyword. You may also have a tool that scores candidates after a video interview, or a chatbot that collects initial responses from applicants. Until recently, all of these were simply efficiency tools. From 2 August 2026, they are classified as high-risk AI systems under European law, and from 2 December 2027 they come bundled with a full set of obligations that Romanian HR teams must be able to demonstrate at any inspection.
The good news: you have not already missed the deadline. The AI Omnibus regulation, Regulation (EU) 2026/1744, signed on 8 July 2026 and published in the Official Journal on 24 July 2026, pushed the full compliance deadline for AI systems listed in Annex III, including those used in recruitment and workforce management, back by 16 months. The less good news: some prohibitions have been active since February 2025, and supervisory authorities have been exercising their powers since August 2026. You have a window until December 2027 to get your house in order, but there is no time to waste.
Below are the four operational changes that directly affect recruitment and evaluation workflows in Romanian companies, with the precise articles from Regulation (EU) 2024/1689 (the EU AI Act) and concrete actions required over the next 16 months.
1. Automated CV filtering and candidate scoring officially classified as high-risk
Annex III, point 4 of Regulation (EU) 2024/1689 classifies as high-risk AI systems any tool intended for "the recruitment or selection of natural persons, notably for placing targeted job advertisements, for analysing and filtering job applications, and for evaluating candidates". The same category covers AI used for task allocation, performance monitoring, promotion decisions, or contract termination.
In practice, this covers: ATS platforms with automatic scoring, AI-powered video interview tools, recruitment chatbots that run screening questions, systems that distribute job advertisements based on candidate profiles, and internal applications that allocate projects or evaluate performance using behavioural data. The European Commission's draft guidance, open for feedback until 23 July 2026, is explicit: a system remains high-risk even when the recruiter makes the final decision, as long as the AI materially influences that decision.
What changes in practice for HR:
- Technical documentation: you must know what data the tool processes, how it builds its recommendation, and what error margins the vendor has disclosed.
- Meaningful human oversight: Article 26(2) requires oversight by persons with "the necessary competence, training and authority." A recruiter clicking "approve" on a pre-scored list of 200 candidates does not meet this standard.
- Logs retained for at least 6 months: deployers must retain automatically generated system logs for a minimum of six months, pursuant to Article 26(6).
- Informing the candidate: Article 26(11) requires that the individual be told they are subject to a high-risk AI system that contributes to a decision about them.
- Informing workers' representatives: Article 26(7) requires that workers' representatives and affected employees be informed before a high-risk AI system is deployed in the workplace.
All of these obligations apply from 2 December 2027, not from 2 August 2026 as originally scheduled. Regulation (EU) 2026/1744 (the AI Omnibus), signed on 8 July 2026 and in force from 27 July 2026, shifted the deadline by exactly 16 months for the areas of maximum exposure: recruitment, workforce management, education, creditworthiness assessment, law enforcement, and critical infrastructure.
2. Emotion analysis in interviews: already banned since February 2025
There is no grace period here. Article 5(1)(f) of Regulation (EU) 2024/1689 prohibits "placing on the market, putting into service for this specific purpose, or using AI systems that infer emotions of a natural person in the areas of the workplace and education institutions, except where the AI system is intended to be put in place or into the market for medical or safety reasons." This prohibition has been in force since 2 February 2025.
In practice: any video interview platform that claims to measure a candidate's "enthusiasm," "sincerity," "self-confidence," or "cultural fit" using vocal inflection, facial expressions, or micro-gestures is outside the law for recruitment purposes. The same applies to any system that analyses the emotional state of employees via office cameras or during client calls.
Exceptions apply strictly for medical or safety purposes, for example, fatigue detection for operators of hazardous machinery, not for evaluating candidates for an account manager role. Analysis by the Future of Privacy Forum notes that the prohibition targets specifically the inference of emotions, not the detection of intent, which falls outside the text of Article 5(1)(f).
The financial exposure here is the highest in the entire Regulation: Article 99(3) provides for fines of up to EUR 35,000,000 or 7% of total worldwide annual turnover, whichever is higher, for violations of Article 5. If you are currently using a vendor that offers this type of emotional analysis, the request to have the feature removed, or the decision to switch supplier, cannot wait.
3. AI literacy: obligation retained, wording softened
Article 4 of the Regulation, also applicable from 2 February 2025, requires providers and deployers to ensure a sufficient level of AI literacy for their staff and for any persons operating AI systems on their behalf. The AI Omnibus changed the verb from "shall take measures to ensure" to the more flexible "shall take measures to support the development of." The distinction matters to lawyers; operationally it changes less: a documented training programme for HR teams using AI tools remains the baseline expectation.
A technology-law specialist quoted in Romanian business media made a point worth internalising: "Systems cannot be left to make decisions on their own, and employees must know how and when to intervene." In any internal audit, the first question a lawyer or authority will ask is who on your team can explain the logic behind the score applied to a rejected candidate. If the answer is "the vendor," you do not have human control.
A second point that dismantles the most persistent market myth: "any company that uses an AI system becomes, in the eyes of the law, a deployer with clear obligations." It does not matter that you did not build the tool. You use it, so you are a deployer, and you fall under Article 26.
4. GDPR, the Romanian Labour Code, and Romanian authorities apply in parallel
The AI Omnibus delay does not suspend the rest of the legal framework. Two obligations remain active for any AI-assisted recruitment process:
- Article 22 GDPR: where a system makes solely automated decisions with a legal or similarly significant effect on an individual, that person has the right not to be subject to such a decision; the controller must ensure transparency, human intervention, and a right to contest. A candidate rejected by an algorithm without meaningful human review can request re-evaluation and a reasoned explanation of the decision.
- DPIA under Article 35 GDPR: Article 26(9) of the AI Act requires deployers to use the information provided by the vendor to carry out a data protection impact assessment (DPIA) wherever one is applicable.
In Romania, the competent supervisory authorities for the AI Act were designated only through a government memorandum dated 12 March 2026: ANCOM (the national communications regulator) as market surveillance authority and single point of contact; ASF (the financial supervisory authority) and BNR (National Bank of Romania) for financial services; ANSPDCP (the national data protection authority) for personal data and sensitive domains; and ADR (the Digital Romania Agency) as notifying authority. The EU deadline for this designation was 2 August 2025, so Romania came in late, but that delay does not suspend the application of the Regulation or exempt companies from their obligations.
What to do concretely in the next 16 months
The window until 2 December 2027 is enough for a realistic plan, not a relaxed one. A compliance roadmap that works for most Romanian companies looks like this:
- Tool inventory: list every AI system used in recruitment, evaluation, task allocation, and monitoring. Include ATS platforms with scoring features, video interview tools, chatbots, targeted job advertisement platforms, and internal performance management systems.
- Risk classification: for each tool, check whether it falls under Annex III, point 4. If it does, request from the vendor the technical documentation required by Article 13.
- Prohibition check: any tool that infers emotions in the workplace must be removed now, not in 16 months. Request written confirmation from the vendor that the functionality has been disabled, or change supplier.
- Human intervention procedure: produce an internal document setting out who, when, and how can override an AI recommendation, with verifiable criteria.
- Candidate notification form: prepare the text by which you inform candidates that an AI system is involved in their assessment, at which stage, and in what capacity.
- DPIA and training: update the data protection impact assessment for recruitment processes with an AI component, and document AI literacy sessions for your HR team.
- Informing workers' representatives: before deploying a high-risk AI system, inform workers' representatives and affected employees as required by Article 26(7).
The bottom line: AI is not banned, the black box is
Regulation (EU) 2024/1689 does not remove AI from recruitment. It removes AI that cannot be explained. The difference between an HR team that walks into December 2027 calmly and one that walks in panicking comes down to three simple things: knowing which tools you use, being able to explain why each one reached a given decision, and being able to tell a candidate that part of their assessment was carried out by an algorithm.
The 16-month extension is not breathing room, it is the budget for documentation you should have been producing anyway. Companies that start today, with the inventory and the removal of already-prohibited features, will arrive in December 2027 with a recruitment workflow that is defensible before ANSPDCP, ANCOM, and any administrative court. Those that delay risk fines of up to EUR 35 million for violations already in force, and operational disruption for those coming into force later.
Read all the articles about
HR Metrics HR Coach Guide Legislative Labour market Workforce planning EthicsArticle written by:
Comments
0 comments
Access your account and add your comment
HR Metrics
Subscribe to the Newsletter
Read articles of interest from Undelucram.ro contributors