The EU AI Act deadline for high-risk hiring AI has moved, and the work it requires is still on the table. Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, amends Article 113 of the AI Act. It gives organizations using Annex III high-risk hiring AI until December 2, 2027 to meet the relevant Chapter III requirements, 16 months later than the original August 2, 2026 date.
The amendment was published in the EU’s Official Journal on July 24, 2026 and entered into force on July 27. For a global capability center (GCC) or IT services firm, the key question for 2027 is simple: can your team show how an AI system influenced a hiring decision for an EU role?
The extra time helps. Some obligations already apply, though, and each runs on its own clock.
Key dates at a glance
| Date | What applies to hiring AI |
|---|---|
| February 2, 2025 | Ban on AI that infers emotions in the workplace (Article 5(1)(f)) |
| August 2, 2026 | Article 50 transparency duties for AI interactions, emotion recognition and deepfakes |
| December 2, 2026 | Article 50(2) marking deadline for synthetic-content systems already on the market before August 2, 2026 |
| December 2, 2027 | High-risk requirements for Annex III systems, including hiring AI |
Annex III, paragraph 4(a) of the AI Act lists AI used to analyze or filter job applications and evaluate candidates as an employment use case. Classification still depends on a use-based assessment under Article 6. A shortlisting or scoring system that materially informs selection deserves close scrutiny under the high-risk rules.
The amended Article 113 delays Chapter III Sections 1 to 3 for Annex III systems until December 2, 2027. These sections cover classification, system requirements, and provider and deployer duties. That includes the Article 26(11) duty to inform people when a high-risk Annex III system makes or assists decisions about them, which follows the same December 2027 timeline.
The amended rules also generally bring high-risk systems already on the EU market before December 2, 2027 into scope only if they are substantially modified after that date. Ask Legal to confirm how this applies to each tool you use, especially if you plan to change vendors or reconfigure workflows.
Article 5(1)(f) has prohibited AI that infers emotions in the workplace since February 2, 2025, and Regulation (EU) 2026/1744 left that ban in place. Recruitment-stage AI that infers enthusiasm, stress or confidence from a candidate’s voice or facial expressions can fall under it. The exception covers systems intended for medical or safety reasons.
A tool that assesses skills or competence raises a separate question, since voice processing alone does not establish emotion inference. Check each system’s purpose and actual outputs before use. A notice or evidence trail added later cannot make a prohibited practice lawful.
A candidate-facing AI interview agent may trigger the provider’s disclosure duty under Article 50(1). Synthetic content may trigger provider marking duties under Article 50(2), and covered deepfakes can trigger deployer disclosure under Article 50(4).
Where an emotion recognition or biometric categorization system may lawfully operate, Article 50(3) requires the deployer to inform the people exposed to it. That notice sits alongside the Article 5 ban and does not override it. Ordinary screening or ranking, on its own, does not create a blanket Article 50 duty to notify candidates.
A team in India can still fall within the Act’s reach. Article 2(1)(c) of Regulation (EU) 2024/1689 covers providers and deployers established outside the EU when AI output is used in the Union. The test focuses on where the output is used, so recruiter location, server location and candidate residency each give only part of the picture.
Consider three common operating patterns:
In each case, trace the output to the EU user and the decision it informs. An EU-established deployer can be another route into scope.
Assign roles at the same time. A firm using an external system for its intended purpose is generally the deployer. The company that develops the system or places it on the market under its own name is generally the provider. Under Article 25, substantially modifying a high-risk system, changing its intended purpose or putting it under your own name can shift provider duties onto your organization. Record who owns configuration, use and documentation. Roles follow what each party actually does, so contract labels alone cannot settle them.
A defensible record lets Legal reconstruct the system, the human judgment, the outcome and the supplier documentation behind a single hiring decision. Build your 2027 AI hiring compliance plan around that standard.
List every point where AI touches hiring: applications, sourcing, assessments, interviews and offers. Give each AI-assisted use its own record, covering:
A scheduling assistant and a ranking system that can remove a candidate from consideration need separate records, even on the same platform.
A final score offers weak oversight if the model has already filtered out applicants nobody reviewed. Define checkpoints before exclusion, at shortlist approval and before offers. Give reviewers the authority to challenge an output, see the relevant inputs and rationale, and restore a candidate to consideration.
When the high-risk duties apply, Article 26 requires deployers to assign human oversight to people with the competence, training and support to carry it out. Test whether reviewers actually use that authority.
For any disputed decision, keep the model or rule version, the relevant output, the reviewer’s identity, the time, the action taken and a short reason for accepting or changing the recommendation. A score alone cannot explain an override. Link each decision to its system record, and avoid copying every resume into a permanent archive.
RippleHire’s controls and visibility page describes its audit trails and role-based access.
Request each vendor’s intended use statement, instructions, known limitations, oversight design, logging capabilities, version change history and available high-risk compliance materials.
Providers and deployers carry different Chapter III duties. The provider builds and documents the system. The deployer uses it within the provider’s instructions, oversees its operation and maintains the records under its control. Build documentation access, incident cooperation and change notification into your procurement terms.
Then run a retrieval drill. Pick one real, access-controlled hiring decision and ask TA, Legal and the vendor to assemble all four records for it. Any gap in that chain shows you where to fix the process.
Evidence needs boundaries. When the high-risk regime applies, Article 26(6) requires deployers to keep automatically generated logs under their control for at least six months, unless other EU or national law says otherwise. Set retention by record type and legal purpose while keeping that minimum where it applies. Limit access to the people investigating decisions, and plan deletion and export around those duties.
The GDPR may add further data protection limits. For teams also handling Indian candidate data, RippleHire’s guide to India’s DPDP Act covers a related workstream. It doesn’t address EU AI Act duties or DPDP commencement dates.
Ownership can be split three ways. The CHRO sponsors the change, the TA head owns checkpoints and records, and Legal validates scope, notices and retention. Together, they’re working toward one outcome: any hiring decision can be explained from the system record.
An evidence chain only works when recruiters can use it inside the workflow they already follow. RippleHire is the ATS where Recruiters and Agents work together. Agents coordinate the workflow and Recruiters stay responsible for selection, while TA leaders and compliance teams share visibility into hiring activity.
Capabilities relevant to an EU AI Act evidence plan include:
Each EU workflow still needs its own check of configuration and legal duties. Start with RippleHire’s explainability and governance information in the Trust Center as you define the documentation you need from any hiring platform.
Book a demo to see how RippleHire’s workflow controls and audit trails can support your EU AI Act recruitment evidence plan.
How should a GCC classify AI that screens applicants for an EU parent?
Start with the tool’s intended use and its effect on selection. A system that filters applications or evaluates candidates fits the employment category in Annex III, subject to the Act’s classification rules. Next, map the GCC, the parent and the vendor against the provider and deployer definitions, and record where the output is used and who acts on it. The parent relationship shapes how the work runs day to day, and classification still depends on the actual workflow.
Which hiring AI interactions need an Article 50 disclosure now?
Look at each specific interaction and piece of content. A candidate-facing AI interview or chat experience may trigger the interaction disclosure rule. Synthetic content can trigger marking or disclosure duties, depending on who produced it and what it is. Where emotion recognition or biometric categorization operates lawfully, the deployer must inform the people exposed to it. Candidate ranking on its own raises a separate question. Document each route that applies and where its notice appears in your hiring process.
What should an IT services firm request from an AI screening vendor?
Ask for intended use instructions, known limitations, logging support, oversight design and a reliable record of model updates. Confirm which records the vendor creates, which ones your firm can export, and how both parties will investigate incidents. If the tool supports EU project staffing, ask how its outputs reach the EU decision maker. Write documentation access and escalation steps into the contract so they work during a real challenge.
What belongs in a defensible recruiter override record?
Capture the recommendation presented, the information the reviewer considered, the action taken and the reason for accepting or overriding it. Link the decision to the tool version and the authorized reviewer. Keep the note short enough for routine use and specific enough that another reviewer can follow it. A good record shows real judgment made at the point of decision.
How can global teams retain hiring evidence without keeping every resume indefinitely?
Separate the evidence needed to explain a decision from the full candidate file. Set purpose-based retention rules for system logs, recruiter reasons, vendor materials and personal data, and apply access controls to each. Legal and privacy teams should reconcile EU requirements with local rules, including DPDP where Indian candidate data is involved. Then test whether a case can still be reconstructed after routine deletion. If it can’t, adjust the record design so the evidence survives while personal data is still deleted on schedule.
This article is for general information and isn’t legal advice. Check your specific obligations with your legal team.