AI Act and employee data: GDPR interplay

EU AI Act 2026: The Complete Compliance Guide for SMEs and Nonprofits

The EU AI Act (Regulation 2024/1689) applies to your SME or nonprofit as soon as you use an AI tool, even a simple ChatGPT subscription. Since February 2025 you must train your teams; on 2 August 2026 the transparency obligations and the penalty regime take effect.

AI Act and employee data: what the overlap with GDPR really changes

Running AI on employee data puts you under two sets of rules that say different things and start on different dates. GDPR and French labour law already govern your HR processing, today, with no grace period. The AI Act adds a further layer, and its heaviest part, the high-risk regime, will not apply until 2 December 2027. In practice, the immediate risk for a small or mid-sized company almost never comes from the AI Act.

That is the point most coverage of this topic misses. Plenty of articles published in 2025 still announce an August 2026 deadline for high-risk AI systems in employment. That date no longer exists. Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, entered into force on 27 July 2026 and pushed the obligations for stand-alone systems listed in Annex III to 2 December 2027. If you built your HR compliance plan on the old timeline, it needs revisiting. For the full picture, our AI Act compliance guide for SMEs covers the whole framework; here we focus on the specific point where AI touches your employees'' data.

What already applies, and what waits until December 2027

Three sets of obligations are in force today, and none of them depends on the high-risk timeline.

The first is GDPR. It never paused. Any employee or candidate data fed into an AI tool remains personal data processing, with a legal basis, a purpose, a retention period, a record of processing activities and data subject rights. Article 22 GDPR additionally governs fully automated decisions, which bears directly on candidate screening.

The second is labour law. In France, three provisions frame any system that observes employee activity: Article L.1121-1 of the Labour Code (proportionality of restrictions on individual freedoms), Article L.1222-4 (informing employees) and Article L.2312-8, II, 4° (informing the works council before deploying new monitoring technologies). These predate AI and apply without waiting for anything.

The third comes from the AI Act itself, but not from the high-risk chapter. Article 4, applicable since 2 February 2025, requires a sufficient level of AI literacy from people operating these systems on your behalf. That includes your HR team. Our article on Article 4 and mandatory AI training sets out what "sufficient level" means in practice. Article 5, on prohibited practices, has also applied since February 2025 and speaks directly to the workplace: emotion recognition at work is banned, except for medical or safety reasons.

For every deadline in sequence, see our complete AI Act timeline.

Which HR uses of AI actually become high-risk

Annex III point 4 does not cover "AI in HR" as a category. It names four precise families of use, and most tools used in smaller companies fall outside them.

High-risk covers systems intended for the recruitment or selection of individuals, in particular to place targeted job advertisements, analyse and filter applications and evaluate candidates. It also covers systems used to decide on promotion or termination, to allocate tasks based on individual behaviour or personality traits, and to monitor or evaluate performance and behaviour within the employment relationship.

The dividing line is the system''s role, not the technology. Asking a conversational assistant to rewrite a job ad produces no assessment of any person, so it is not a high-risk use. Having the same tool summarise interview notes the recruiter wrote themselves is not either. Plugging in a tool that assigns a compatibility score to each CV and returns a ranking, however, brings you inside Annex III, even if a human signs off afterwards. The text refers to systems intended to filter and evaluate candidates, without requiring that the final decision be automated. Our article on high-risk AI systems works through the qualification criteria.

Three cases that cause genuine hesitation:

• An ATS that sorts on keywords with no AI model: not an AI system under the Regulation, so no Annex III qualification. GDPR and Article 22 still apply in full if the sort discards applications without human involvement.
• A productivity measurement tool: software that counts files handled and reports it transparently to the manager on a quarterly basis is treated by the French data protection authority as proportionate. The same count in real time, employee by employee, tips into constant monitoring and becomes excessive.
• A general-purpose AI assistant used by a manager to draft an annual review: the system is not doing the evaluating, the manager is. But if the review content is generated from performance data imported into the tool, you have created undeclared processing and, potentially, an evaluation use within the meaning of Annex III.

Employee data: the real trap sits upstream of the AI Act

In most of the companies we audit, the problem is not high-risk qualification. It is HR data flowing into consumer-grade tools.

An HR director who pastes a compensation table into a free AI assistant to get an analysis is transferring personal data, sometimes special category data, to a processor with no contract, often outside the European Union, with no identified legal basis and no trace in the record of processing activities. That is not an AI Act issue. It is a GDPR issue, and it is enforceable today.

The concrete checkpoints, in the order we work through them:

1. The account in use. A business account with a contractual commitment not to train on your data, or nothing. Consumer plans generally default to the opposite.
2. Location and transfers. Hosting, sub-processors, transfer mechanism outside the EU. To be documented, not assumed.
3. The data perimeter. No health data, no trade union membership, no ethnic origin, no disciplinary records in a prompt. These categories fall under Article 9 GDPR and their processing is prohibited as a matter of principle.
4. Entry in the record. Any recurring use of an AI tool on HR data is processing and belongs in the record. Our method for mapping your AI systems includes the template we use on assignments.
5. The impact assessment. A GDPR data protection impact assessment becomes necessary as soon as automated processing evaluates people at meaningful scale, which is nearly always true of a screening tool.

Shadow AI is central here. A written, circulated usage policy costs little and resolves a large share of the exposure: we publish a reusable AI policy template for businesses.

Informing employees and the works council: two obligations that stack

The AI Act and labour law each impose prior notice, with different scopes and different timelines. One does not replace the other.

On the AI Act side, Article 26(7) provides that before putting into service or using a high-risk AI system at the workplace, deployers who are employers shall inform workers'' representatives and the affected workers that they will be subject to its use. This obligation follows the high-risk timeline, so December 2027 for Annex III uses.

On the labour law side, the obligation is immediate and broader. According to the French data protection authority, a system monitoring staff activity must meet three cumulative conditions: pass the justification and proportionality test, be submitted to employee representative bodies under the applicable rules, and be brought to employees'' attention before deployment. Works council consultation is mandatory in companies with 50 or more employees, prior to implementation. The authority cites as unlawful the installation of a time clock or video surveillance system without prior works council consultation in a company of that size. The reasoning is identical for an AI tool that monitors activity.

On form, a note read out in a meeting will not prove compliance. The employer must be able to demonstrate that these conditions were met: the necessity and proportionality analysis in detail, the data lifecycle, and the measures taken for information and the exercise of rights. Our article on the AI information notice sets out the expected structure and the wording not to omit.

Automated decisions: exactly where the red line sits

Article 22 GDPR prohibits, as a matter of principle, decisions producing legal effects or significantly affecting a person being taken solely on the basis of automated processing.

Applied to recruitment, that means a screening tool cannot reject an application on its own. Human involvement has to be real, not formal. A recruiter who approves a batch of rejections produced by a score without looking at the files is not providing meaningful human involvement: the decision remains automated in substance. GDPR further requires that the data subject be able to obtain human intervention, express their view and contest the decision.

Three operational consequences:

• A score threshold must not trigger automatic rejection. It can order a pile, not empty it.
• The recruiter needs a way to understand why a candidate ranks poorly, otherwise the review is decorative. This is also what the AI Act will call effective human oversight from December 2027.
• The contestation route must exist and be communicated to candidates, not merely recorded in an internal policy.

For how the two texts fit together, our comparison of the AI Act and GDPR sets out the difference in logic: GDPR protects the person and their data, the AI Act governs the product and its lifecycle.

The company with no DPO and no works council: who does what

A thirty-person company has no mandatory data protection officer, no works council and no in-house lawyer. That is not an exemption, it is a different allocation of roles.

What we put in place in that situation, and what it takes a few days of work:

• A named owner. One person, usually the managing director or the administrative manager, who carries the topic and holds the record. With no mandatory DPO, the controller''s liability remains entirely intact.
• An inventory of HR uses of AI. A ten-row table is enough to start: tool, purpose, data involved, who uses it, business account or not.
• A one-page usage policy. What is allowed, what is not, which account to use, which data never to enter.
• Written notice to employees. Even without a works council, Article L.1222-4 of the Labour Code applies: employees must be informed of collection systems that concern them.
• A short training session. Article 4 of the AI Act targets your team, not your vendor. A well-scoped half-day covers most of the usage risk.

Our AI Act guide for small businesses lays out this lighter path, and our article on the conformity assessment describes the real steps when you have no legal department.

Penalties and enforcement: what actually threatens you in 2026

Today''s enforcement risk sits with GDPR and labour law, not with the AI Act''s high-risk obligations, which are not yet applicable.

GDPR provides for administrative fines of up to 20 million euros or 4 % of worldwide annual turnover. In practice, amounts imposed on French small and mid-sized companies run in the tens of thousands of euros, but the regulator increasingly uses its simplified sanction procedure: it issued 23 new sanctions through that route between January and early July 2026.

More to the point, recruitment is one of the French regulator''s priority enforcement themes for 2026. Around 20 % of its annual inspections fall under these themes, and this one explicitly targets automated decision-making systems, candidate information and retention periods. The authority says it will focus first on large companies and recruitment agencies, but it also states that the campaign foreshadows its future role as market surveillance authority in the employment field under the AI Regulation. In other words, the practices it documents this year will shape the doctrine that applies to everyone afterwards.

On the AI Act side, Article 5 on prohibited practices is already enforceable at up to 35 million euros or 7 % of worldwide turnover. It is the only part of the Regulation exposing a smaller company to a fine today on an HR topic, through the ban on emotion recognition at work. Other obligations, including those of deployers, are capped at 15 million euros or 3 % and follow the December 2027 timeline. Our article on AI Act penalties breaks down the tiers by type of breach.

Frequently asked questions

Can we use ChatGPT or a comparable AI assistant for HR tasks?

Yes, with a business account that guarantees no training on your data and a defined data perimeter. Drafting a job ad, preparing an interview grid or rewriting an internal memo raises no difficulty. Analysing real CVs, named performance reviews or compensation data in a consumer tool exposes you to a GDPR breach. Our article on ChatGPT and AI Act compliance covers the configuration in detail.

Do we have to consult the works council before deploying an AI tool in HR?

In a company with 50 or more employees, yes, if the tool constitutes an activity monitoring system or a new technology affecting working conditions. Article L.2312-8, II, 4° of the French Labour Code requires informing the works council before new monitoring technologies are deployed. Below 50 employees, works council consultation does not apply, but individual notice to employees under Article L.1222-4 is still owed.

Is a CV scoring tool banned?

No. It is classified as high-risk by Annex III point 4, which triggers documentation, human oversight and log retention obligations from 2 December 2027. What is prohibited today is letting that score decide on its own to reject an application: Article 22 GDPR stands in the way.

Is emotion recognition in interviews legal?

Not at the workplace. Article 5 of the AI Act bans emotion inference systems in the workplace and in education, except for medical or safety reasons. The ban has applied since 2 February 2025, and a recruitment interview for a role at the company falls within that scope.

How long should we keep CVs analysed by AI?

Using AI does not extend the retention period. It stays aligned with the purpose of the processing, meaning two years after the last contact for a speculative application or a talent pool under French regulatory guidance, provided the candidate has been informed. Retention periods are among the points the regulator is explicitly checking in 2026 on recruitment.

What if our payroll provider uses AI without telling us?

You remain the controller for your employees'' data. Review the processing agreement, obtain a description of the automated processing performed and, where needed, have the technical annex updated. A provider who refuses to document its AI usage is a risk point to record in writing.

Do we have to train the whole company or just the HR team?

Article 4 of the AI Act targets people operating AI systems on the company''s behalf. If only three people in HR use these tools, the obligation covers them. In practice usage spreads quickly, and we recommend covering the functions that handle personal data first: HR, sales, management.

How GrowthPerf supports smaller companies on this

We work on this overlap of AI Act, GDPR and labour law in a short format, calibrated for organisations with no DPO and no legal department.

The HR audit starts with an inventory of actual AI usage, including the uses management does not know about. We then qualify each one: out of scope, GDPR processing to be brought into order, or future high-risk under Annex III. You come out with an up-to-date record, a policy you can circulate, the notice documents for employees and the works council, and a prioritised list of what has to be fixed before December 2027.

Training completes the picture. Our AI for business programme covers the Article 4 obligation and the reflexes for handling personal data, and the operational AI track goes further for teams building their own use cases. GrowthPerf holds the French Qualiopi certification, which opens up funding through your sector training fund.

The most effective starting point is a status check. A free 30-minute audit is enough to tell whether your HR uses of AI are already in breach or simply undocumented. For the full framework, go back to our AI Act compliance guide for SMEs and non-profits, which lists every obligation by deadline.