EU AI Act HR Deadline Delayed to 2027: What to Do Now

Marta, the HR manager at a mid-sized Baltic company, has spent several weeks building a folder titled "AI Act – 2 August." It holds a list of the AI tools the company uses for CV screening and performance evaluation, a draft outline of who should be assigned to human oversight, and notes on what data the systems ought to be logging. She has been preparing for a deadline that was supposed to land on 2 August 2026.

Then a colleague sends her a link to a legal-news story: the deadline for high-risk employment AI obligations has been pushed back by 16 months. Marta's first reaction is relief. Her second is confusion. Does the whole folder go into a drawer now? Is there still anything due in August? Most of the articles she has read so far — including some legal-portal overviews — were written before the change and still repeat the old 2 August deadline as if it applied across the board.

This article separates two things: what is still mandatory for employers on 2 August 2026, and what has actually been pushed to 2 December 2027. At the end, there's a short list of what's worth doing now regardless of the delay.

What actually changed on 29 June

On 16 June 2026, the European Parliament approved the so-called AI Omnibus — the European Commission's proposal to make certain AI Act obligations easier to implement and their timeline more flexible. On 29 June, the Council gave it the final green light too. The regulation is expected to be published in the EU Official Journal in July 2026 and enters into force three days after publication — so in practice, the change is already here or right around the corner.

The change that matters most for HR: the application date for Annex III stand-alone high-risk AI systems — including AI used for recruitment, candidate selection, performance evaluation, task allocation, worker monitoring, and promotion or termination decisions — moved from 2 August 2026 to 2 December 2027. That's 16 extra months for exactly the heavy documentation burden that most of the legal commentary published so far has been about. On top of that, the deadline for Annex I embedded high-risk systems — such as certain AI components built into products — moved to 2 August 2028. That one touches the HR context less often, but it belongs to the same package.

It's important not to read this as the whole AI Act being put on hold. The Omnibus changed deadlines and some administrative requirements, not the principles of the regulation. Three blocks of obligations that HR managers need to know about right now were left completely untouched by the Omnibus — and these are exactly the ones that articles written before 29 June mostly under-emphasize, because at the time of writing they had either only recently taken effect or were still on their way.

(One theoretical caveat: if the Omnibus somehow failed to enter into force before 2 August 2026, the original AI Act timeline would apply exactly as first written. Every source treats formal adoption as already complete, though, so this is a theoretical scenario rather than a realistic one.)

Two ways HR managers are getting this wrong

Marta's situation illustrates two possible mistakes, and they carry very different weight.

The first mistake is over-preparing for a deadline that no longer exists. If the HR department spends the summer building documents — a fundamental rights impact assessment, database registration prep, a six-month log architecture — whose deadline is now December 2027, that's not a disaster, but it is time and attention that could have gone elsewhere. This mistake is costly, not dangerous.

The second mistake is more dangerous: assuming that because the deadline moved, nothing at all applies to employers in August 2026 anymore. That's wrong, and it leaves real non-compliance risk uncovered. The emotion-recognition ban, the AI literacy obligation, and most of the transparency duties are not tied to the Annex III high-risk classification that the Omnibus delayed — they come from separate articles with their own, unchanged timelines.

Three buckets: what applies, what's delayed, what's worth doing anyway

The most practical way to look at the situation is to sort the obligations into three buckets.

Already in force today, deadline unchanged:

• The ban on emotion-recognition AI in the workplace (Article 5) — in force since February 2025. This is not a "manage the risk" category; it's an absolute prohibition: an employer may not use an AI system that infers the emotions of employees or candidates, for example by analyzing facial expressions or tone of voice.
• The AI literacy obligation (Article 4) — also in force since February 2025. The employer must ensure that staff who use AI systems, or whose work is affected by them, receive adequate training, and this should be documented, not just verbal guidance.
• Most of the transparency duties (Article 50) — take effect on 2 August 2026, exactly as originally planned. This includes the duty to tell people they're interacting with an AI system, for example an AI chatbot in a recruitment process, and the duty to disclose when emotion-recognition or biometric categorization AI is used on someone. Only the Article 50(2) watermarking duty (labelling AI-generated content) got a narrow deferral to 2 December 2026, and even that only for systems already on the market before 2 August 2026.
• General-purpose AI model (GPAI) provider obligations — in force since August 2025, untouched by the Omnibus.

Bought you time — deadline now 2 December 2027:

• Assigning competent human oversight to a high-risk employment AI system (Article 26(2)).
• The Fundamental Rights Impact Assessment (FRIA), where required (Article 26(9)).
• Registering the system in the EU high-risk AI database (Article 26(8)).
• The minimum six-month log-retention obligation (Article 26(6)).
• The full worker and worker-representative notification duty tied to the high-risk classification (Article 26(7)).

Worth doing anyway, because it reduces risk ahead of the 2027 deadline:

• Assign someone now who is responsible for reviewing AI recommendations, even if formal documentation isn't required yet — in 17 months, that person will already be experienced in the role, not freshly appointed.
• Collect a written or digitally signed acknowledgment from staff confirming they've read the company's AI use policy. It's a cheap, quick step that creates a documented trail today, instead of scrambling at the end of 2027 to reconstruct who knew what and when.

Example: a CV-screening AI tool against the three buckets

Suppose a company uses an AI tool that screens incoming CVs and ranks candidates by fit. How does that sort into the three buckets?

Candidate screening and selection falls under the AI Act's Annex III high-risk category — that hasn't changed. If the tool only analyzes CV content (work experience, skills, education) and doesn't detect emotions, it doesn't violate Article 5. But if the tool also has a video-interview analysis feature that assesses a candidate's facial expressions or tone of voice to detect stress or motivation, that's a prohibited practice already today — regardless of what happens in 2027.

Two things need to happen right now, regardless of the delay: recruiters who use the tool need documented training on how it works and what its limitations are (Article 4); and candidates need to be told when an AI system is part of the process, for example at the pre-screening stage (Article 50(1)).

Only the part about formal documentation — a named human overseer, the impact assessment, database registration, six-month logs — is the part whose deadline moved to 2 December 2027. A company that has already assigned a recruitment manager to review every AI-generated ranking before a decision is made doesn't need to redo that practice by 2027 — it simply formalizes what's already in place once the deadline arrives.

How to document this cheaply, starting today

The simplest step most companies can take this month isn't technical — it's paperwork. If the company already has a short AI use policy document (what is and isn't allowed with AI tools, who is responsible for oversight), the most practical move is to give it to staff for signature and archive the acknowledgment.

This is where Agrello has a natural, if limited, role. A company can build a single policy document template with variable fields (name, department, date) and send it out for signature to the whole team, or to the specific department using AI tools, all at once. Each employee's acknowledgment gets signed and lands automatically in an archive with a full signing history, or audit trail — who read and confirmed the document, and when. This isn't AI Act compliance proof, and it doesn't replace the formal documentation required by 2027, but it's exactly the kind of paper trail that later compliance will require anyway, and building it now costs a fraction of what assembling the same thing in a rush at the end of 2027 would.

Bottom line

The AI Act deadline for high-risk employment AI documentation moved from 2 August 2026 to 2 December 2027 — that's a fact, and Marta's folder can go ahead and set that part aside. But standing right next to that folder are three obligations that haven't gone anywhere: the emotion-recognition ban, the AI literacy requirement, and most of the transparency duties apply from exactly the date originally planned. Anyone who confuses the two either wastes time on a deadline that's already been pushed back, or leaves a real August obligation unmet.

If you want to start formalizing your AI use policy document and collecting signed acknowledgments from staff today, log in to Agrello or create a new account and send the first template out for signature today.