Skip to main content
Employer Guides

EU AI Act Résumé-Screening Deadline: Now 2 December 2027

Calin Muresan•
#EU AI Act#AI in recruitment#résumé screening#Digital Omnibus#HR compliance#Romania

EU AI Act Résumé-Screening Deadline: Now 2 December 2027

By Calin Muresan, co-founder of Wise Step Recruiting. Calin spent about 20 years as a software engineer and cloud architect before he started recruiting engineers.

If you’ve read that EU AI Act enforcement against résumé-screening tools “has started,” the law says otherwise. The Digital Omnibus on AI is Regulation (EU) 2026/1744, an amending act that simplifies and re-times the AI Act. It entered into force on 27 July 2026. It moved the compliance date for stand-alone high-risk AI systems — the category that covers CV ranking and candidate filtering — from 2 August 2026 to 2 December 2027. That is sixteen months of extra runway, written into the Official Journal.

It is not a free pass. Some rules that hit hiring tools have applied since February 2025, and the fine ceilings for large companies did not move. Below: what the regulation changed, what it left alone, where the “enforcement has started” story came from, and what to do this quarter if you run or are buying AI screening. For the full obligations map, see our EU AI Act guide for recruitment teams.

Key Takeaways

  • Regulation (EU) 2026/1744 sets 2 December 2027 as the date high-risk obligations apply to Annex III systems, including tools that “analyse and filter job applications” (EUR-Lex).
  • Fine ceilings stay at €15M or 3% of global turnover for high-risk breaches and €35M or 7% for prohibited practices. The only change: small mid-caps now get the “whichever is lower” cap SMEs already had, on the lower two tiers.
  • The ban on inferring emotions from biometric data at work has applied since 2 February 2025. Commission guidance reads it as covering job candidates — which, on our reading, includes video-interview analysis of face and voice.
  • Systems already on the market before 2 December 2027 generally fall in scope only after “significant changes in their designs.” For a SaaS screener whose vendor retrains or swaps its ranking model, that exemption may not hold.

What Did Regulation (EU) 2026/1744 Change?

It deferred the high-risk rulebook. The amended Article 113 of the AI Act now says Chapter III, Sections 1, 2 and 3 apply from 2 December 2027 for systems classified high-risk under Annex III. For AI embedded in regulated products under Annex I, the date is 2 August 2028 (Regulation (EU) 2026/1744, Art. 1(40), read on EUR-Lex 2026-10-05). Those sections hold the classification rules, the requirements — risk management, data governance, documentation, human oversight — and the provider and deployer obligations.

The dates are on the face of the act. Parliament and Council signed it in Strasbourg on 8 July 2026. It appeared in the Official Journal on 24 July and entered into force “on the third day following” publication — 27 July 2026 (Regulation 2026/1744, Art. 4; European Commission). Gibson Dunn and DLA Piper read the deferral the same way at the agreement and proposal stages.

The regulation gives its own reason. Recital 40 cites “the delayed availability of standards, common specifications, and alternative guidance” and the delayed establishment of national competent authorities. In plain terms: providers had no harmonised standards to demonstrate conformity against, and the authorities meant to supervise them were not all in place.

Recruitment sits squarely inside the deferred group. Annex III, point 4(a) names AI systems used “to analyse and filter job applications, and to evaluate candidates” (consolidated AI Act, 27.07.2026). The Article 6(3) carve-out for low-impact Annex III tools does not help here: an Annex III system “shall always be considered to be high-risk” when it profiles people, and CV ranking typically does, because it evaluates candidates’ personal aspects (GDPR Art. 4(4)). If your ATS ranks CVs or scores candidates, this is your date. For what that means on the hiring side, see AI and data science recruitment in Romania.

EU AI Act application dates before and after Regulation (EU) 2026/1744 Timeline from 2025 to 2028. Prohibitions and AI literacy in force since 2 February 2025. Annex III high-risk, including résumé screening, moved from 2 August 2026 to 2 December 2027, a 16-month delay. Annex I product-embedded AI moved from 2 August 2027 to 2 August 2028, a 12-month delay. What moved, what didn't 2025202620272028 Prohibitions + AI literacy in force since 2 Feb 2025 Annex III high-risk incl. CV screening 2 Aug 2026 2 Dec 2027 (+16 mo) Annex I (in products) 2 Aug 2027 2 Aug 2028 (+12 mo)
Source: Regulation (EU) 2026/1744, Art. 1(40), amending Art. 113 of Regulation (EU) 2024/1689 (EUR-Lex, retrieved 2026-10-05). Positions plotted to scale.

What Did Not Change — and Already Applies Today

Three things matter for hiring teams right now, whatever the new date.

The emotion-recognition ban. Article 5(1)(f) prohibits AI that infers emotions “in the areas of workplace and education institutions,” except for medical or safety reasons. It has applied since 2 February 2025. The Act defines an emotion recognition system as one working “on the basis of their biometric data” — face, voice, physiological signals (Art. 3(39)). The Commission’s Guidelines on prohibited AI practices (4 February 2025) read “workplace” as covering candidates during selection.

The Guidelines are non-binding — only the Court of Justice gives authoritative interpretations — but they show how authorities are likely to read the ban (Future of Privacy Forum analysis). A video-interview tool that scores “enthusiasm” or stress from a candidate’s face or voice is not waiting for 2027.

AI literacy. The Omnibus rewrote Article 4 but kept the duty. Providers and deployers must “take measures to support the development of AI literacy” of the staff operating their AI systems. The new text adds that this “does not require” guaranteeing “any specific level of AI literacy of any individual” (Regulation 2026/1744, Art. 1(5), replacing Art. 4). A recruiter who cannot say what the tool’s score measures is the kind of gap this duty is meant to close.

The fine ceilings. Article 99 still caps prohibited-practice fines at €35M or 7% of worldwide annual turnover, and most other operator breaches — high-risk included — at €15M or 3%, whichever is higher (consolidated AI Act, Art. 99). SMEs already paid whichever amount is lower, on every tier (Art. 99(6)). The Omnibus adds paragraph 6a, which gives small mid-caps (SMCs) the same “whichever is lower” rule for the €15M and €7.5M tiers only. These are EU ceilings. Each member state sets the actual penalty rules (Art. 99(1)), which is why national implementing laws matter.

Breach EU ceiling (higher of) SMEs SMCs
Prohibited practices (Art. 5) €35M or 7% of turnover lower of the two higher of the two
High-risk and most other operator duties €15M or 3% lower of the two lower of the two (new)
Incorrect or misleading information to authorities €7.5M or 1% lower of the two lower of the two (new)

GDPR never moved either. Article 22 gives candidates the right not to be subject to a decision based solely on automated processing that significantly affects them. Recital 71 names “e-recruiting practices without any human intervention” as an example (GDPR, EUR-Lex). Fully automated rejection was a data-protection problem before the AI Act and still is. How stable are automated screening scores in the first place? See AI résumé screening: same CV, 100 runs, scores 66 to 99.

EU AI Act maximum fine ceilings after the Digital Omnibus Lollipop chart of maximum fines. Prohibited practices: 35 million euros or 7 percent of global turnover; SMEs pay the lower amount, SMCs do not. High-risk and most other operator obligations: 15 million euros or 3 percent. Incorrect information to authorities: 7.5 million euros or 1 percent. On these two tiers both SMEs and SMCs pay the lower amount. Maximum fine ceilings after the Omnibus Prohibited practices (Art. 5) €35M or 7% · SME: lower of · no SMC cap High-risk & other duties €15M or 3% · SME/SMC: lower of Misleading info to authorities €7.5M or 1% · SME/SMC: lower of Line length scaled to € ceiling
Source: Regulation (EU) 2024/1689, Art. 99(3)–(6a), consolidated 27.07.2026 (EUR-Lex, retrieved 2026-10-05).

Where Did “Enforcement Has Started” Come From?

From secondary sites, not from Brussels. Our daily market briefing re-checked one claim — that the European AI Office named automated résumé screening in a “first inspection wave” — in seven consecutive editions between 25 September and 1 October 2026. Each check traced the claim to its origin, then looked for a matching Commission press release, AI Office decision, or Official Journal entry. The claim appeared on at least six compliance-marketing and aggregator sites. None cited a primary EU source, and we found none.

The claim fails on its own logic. No authority can inspect a provider for breaching high-risk obligations that do not apply until 2 December 2027. It also names the wrong body for most tools. Under Article 74, national market surveillance authorities supervise Annex III systems. The AI Office’s direct enforcement powers centre on general-purpose AI models (Arts. 88–94). Since the Omnibus, they also reach AI systems a provider builds on its own general-purpose model. A typical third-party screening tool answers to the national authority.

What actually happened is narrower. On 1 September 2026, the Commission confirmed it had sent its first requests for information to more than 30 AI providers. The requests covered model safety and security, plus copyright and transparency (Agence Europe). Those are general-purpose AI obligations. Real enforcement, different target.

In Romania, the national side is further behind. A government memorandum of 12 March 2026 designated ANCOM as market surveillance authority and single point of contact. That came seven months after the August 2025 EU deadline (Mararu). As Juridice.ro notes, a memorandum does not itself grant sanctioning powers: as of 13 August 2026, the national law setting control procedures and sanctions was still being drafted (Juridice.ro). We found no report of its adoption as of 5 October 2026. More regional context: IT recruitment trends in Eastern Europe, 2026.

The Grandfathering Clause Is Narrower Than It Looks

Article 111(2), as rewritten by the Omnibus, says the AI Act applies to high-risk systems placed on the market or put into service before the Chapter III date “only if, as from that date, those systems are subject to significant changes in their designs” (Regulation 2026/1744, Art. 1(39)). Systems used by public authorities are the exception: they must comply by 2 August 2030 regardless.

Read fast, that sounds like a reason to sign a screening contract in 2026 and forget about it. Read as an engineer, it is weaker. The Act does not define “significant change in design.” It is also untested whether a hosted SaaS service counts as one system placed on the market once, or as something re-released with each update. Our reading: if your vendor retrains its ranking model, swaps the underlying LLM, or changes the scoring features after 2 December 2027, it will struggle to argue the design has not significantly changed. Ask your vendor how often it does each of those. The exemption protects frozen systems.

The clause also does not touch Article 5 or GDPR. A grandfathered screener that infers emotion from biometric data, or auto-rejects without human review, is exposed today.

What Should Hiring Teams Do Now?

Use the sixteen months; don’t spend them waiting. In order of urgency:

  1. This week — audit for prohibited features. List every AI tool in your hiring funnel. Switch off anything that infers emotion from face, voice, or other biometric data. That is the €35M / 7% tier, live since 2025. Review personality-scoring features separately: they are not banned, but they are Annex III high-risk and a GDPR Article 22 risk.
  2. This month — put a meaningful human on every rejection. Map where your ATS auto-rejects or auto-ranks. To stay outside Article 22, each decision that ends a candidacy should have a reviewer with the authority and information to change the outcome — not a name on a rubber stamp (GDPR Art. 22, Recital 71).
  3. This quarter — document AI-literacy measures. Article 4 already applies. Record who uses each tool, what training they had, and when.
  4. This quarter — ask vendors for their 2027 plan in writing. Request their Annex III classification, technical documentation roadmap, and how they define “significant change” in their release process. A vendor with no answer now is unlikely to be ready by late 2027.
  5. Before 2 December 2027 — prepare deployer obligations. Human oversight, log retention, input-data checks, and informing candidates. These are deployer duties; your vendor cannot carry them for you. Our EU AI Act compliance checklist for recruitment breaks them down item by item.

Do not:

  • Rip out a working ATS on the strength of an “enforcement has started” headline.
  • Treat the new date as permission to deploy emotion-scoring interview tools.
  • Assume grandfathering covers a SaaS product whose model changes after the deadline.

Why the Deadline Moved — and What to Watch Next

The date itself is fixed in Article 113. Whether providers can meet it depends on harmonised standards from CEN-CENELEC landing in time. Standards are not mandatory, but they give providers a presumption of conformity (Art. 40). Without them, providers build toward the requirements with no agreed way to show they meet them — the same problem Recital 40 cites for this delay.

Watch two things: CEN-CENELEC standard publications through 2027, and Romania’s implementing law. For hiring managers, the practical point is older than the AI Act. A tool that ranks engineers on keyword density was never a good screen. The regulation puts a date on documenting what a screen measures and how accurately (Arts. 10 and 15). How we screen instead: AI-assisted technical screening, with engineers in the loop.

Frequently Asked Questions

Is EU AI Act enforcement against résumé-screening tools in force?

No. The high-risk obligations covering CV screening apply from 2 December 2027 under Regulation (EU) 2026/1744. What applies today is the ban on inferring emotions from biometric data at work, which Commission guidance reads as covering candidates, and the AI literacy duty. Both have applied since 2 February 2025.

Did the Digital Omnibus reduce AI Act fines?

No. The ceilings remain €35M or 7% of global turnover for prohibited practices and €15M or 3% for high-risk and most other breaches. The only change lets small mid-caps pay the lower of the two amounts on the €15M and €7.5M tiers. SMEs already paid the lower amount on all tiers.

Does a résumé-screening tool bought in 2026 need to comply in 2027?

Generally only if its design changes significantly from 2 December 2027, under the amended Article 111(2). Systems used by public authorities must comply by 2 August 2030 regardless. “Significant change” is undefined; on our reading, SaaS tools that retrain or swap ranking models are likely to meet it. Five questions to put to any vendor are in our AI résumé screening reliability test.

Who enforces the AI Act on hiring tools in Romania?

ANCOM, designated by government memorandum on 12 March 2026 as market surveillance authority and single point of contact. As of August 2026, Romania had not adopted the national law that sets AI Act sanctions, and we found no report of its adoption as of 5 October 2026.

References and Retrieval Notes


Last updated: October 5, 2026. Sources retrieved October 5, 2026; see the references above.