Skip to main content
Podcast

Five Steps for Amended EU AI Act Compliance

Details

November 17, 2026

The EU Digital Omnibus Package has changed the timeline for certain EU AI Act obligations, but not their overall direction. In this episode, L+E Global colleagues Luís Neto Galvão and Éric Le Quellenec offer five practical steps — inventory and classify AI-enabled HR tools, strengthen oversight and vendor contracts, mitigate bias, train relevant teams, and prepare recruitment practices for compliance — that GCs and HR leaders at multinational businesses should take now to fortify their employment decision-making AI programs. 

Transcript

Luís Neto Galvão
Partner, SRS Legal

Today, we're talking about artificial intelligence in human resources and about a legal text that has just changed the timetable: the Digital Omnibus. Before we get into the practical side, can you remind us, Eric, what exactly happened and what has and hasn't been postponed?

Éric Le Quellenec
Partner, Flichy Grangé Avocats

Thank you, Luís. Happy to be here with you. Back in November 2025, the Commission presented the Digital Omnibus, a package to simplify European Union (EU) digital law. It has two parts. The first reopens the General Data Protection Regulation (GDPR). It is still being negotiated, so we'll leave it aside today. The second amends the AI Act, and that one is now law. It entered into force on July 27, 2026. The headline change is the delay for high-risk systems.

The obligations for the systems listed in Annex III, including recruitment and career management, were due to apply in August of this year. They now apply from December 2, 2027, and from August 2, 2028, for AI built into regulated products.

Not everything has been postponed. The prohibitions have applied since February 2025, the AI literacy obligations as well, and the transparency rules since August 2026. Clients should not think they have nothing to do until 2027.

Neto Galvão

That's really our message today, Eric. If our clients haven't started, now is the time to put a compliance plan in place. A year goes very quickly when you have to map your tools, renegotiate vendor contracts, and train your teams. Let's go through the main steps with concrete examples.

Regarding the obligation to map and classify, the first step is obviously to know what you have. Where should an employer start, and how does it work out whether the tools are actually caught by the AI Act?

Employers should have an advantage because they often forget that their GDPR record of data processing activities is a very useful tool. Recruitment, performance reviews, training, and internal mobility are all already listed, and that is the natural starting point.

Then, for each line, there are two questions: Is there AI involved? If so, what does it actually do?

Le Quellenec

On the first question, the key criterion is inference. The system infers from its inputs how to produce outputs, predictions, recommendations, or decisions, rather than simply applying rules someone wrote. A spreadsheet that filters candidates by years of experience is not AI. A tool that learns from past hiring decisions to rank applicants is.

Then comes the risk level, and three levels matter for HR. First, prohibited practices under the AI Act. For employers, the key one is emotion recognition in the workplace, save for medical or safety reasons. A video interview that scores a candidate's enthusiasm from their face is not high risk; it is simply prohibited. Some of those tools are still on the market.

Second, high-risk systems. Point 4 of Annex III covers two families: recruitment, including targeting job ads, filtering applications, and evaluating candidates; and decisions on the employment relationship, including promotion, termination, task allocation based on behavior or personal traits, and monitoring and evaluating performance.

Third, the exception. A listed system is not high risk if it performs a narrow procedural task, improves the result of an activity a human has already completed, detects patterns without replacing human assessment, or performs a preparatory task. But there is a firm limit: a system that profiles people is always high risk.

Neto Galvão

To make that concrete for our listeners, a tool that extracts data from CVs into the applicant database is preparatory, so it is probably outside the scope. A tool that checks a job ad for gendered wording after HR has drafted it is also outside. But a tool that scores and ranks candidates is high risk, without a doubt.

The same logic applies beyond hiring. A tool that scores employees' performance, recommends who should be promoted, or flags who is likely to resign is profiling. It is high risk, whatever the vendor says.

I would add one practical point. When a client relies on the exception, it must document that assessment beforehand. That is the first document the regulator, or the other side in litigation, will ask for.

Another practical warning: the official procurement list is rarely the full picture. The employer should also look for what is called shadow AI, meaning tools adopted directly by recruiters, managers, or local teams, including AI functions quietly added to existing HR software. The inventory should therefore cover both centrally approved systems and systems actually in use on the ground.

Now, in relation to governance, once the high-risk tools are identified, the question becomes who is responsible for what. What does the employer concretely have to do, Eric?

Le Quellenec

Usually, the vendor is the provider and the employer is the deployer. But this can change. If the employer puts its name on the tool, modifies it substantially, or changes its intended purpose so that it becomes high risk, it becomes a provider, which brings much heavier obligations.

Before buying, the employer should check what the provider has done: the conformity assessment, CE marking, EU declaration of conformity, and registration in the EU database. That is due diligence, and it belongs in the contract.

As a deployer, the rules are very concrete. Use the system in line with the instructions. Assign human oversight to people with the competence, training, and authority to overrule it. Make sure the input data the employer controls are relevant and representative. Monitor the system, and if a risk appears, suspend use immediately and inform the provider. Serious incidents must be reported. Keep the logs for at least six months.

Use the provider information to carry out the GDPR data protection impact assessment (DPIA). Inform candidates and employees who are subject to decisions. Before the tool is introduced, inform workers' representatives and the workers concerned.

On top of that, individuals have a right to an explanation of the role AI played in a decision that significantly affects them. A rejected candidate can ask for it.

One common misconception concerns the fundamental rights impact assessment introduced by the AI Act. It applies to public bodies, private entities providing public services, and credit and insurance uses. A private employer generally does not have to do it. For them, the key assessment remains the GDPR DPIA.

Neto Galvão

I would add that human oversight must be designed into the process, not added as a label. The employer should define which decisions the system may support, which decisions remain exclusively human, when a recommendation must be challenged, and who has the authority to stop using the tool.

A manager who can only confirm the system's output is not providing meaningful oversight. In the employment context, implementation is also an industrial relations issue. Before introducing a high-risk system at work, the employer should identify any applicable information or consultation obligations toward workers' representatives. The employer should explain, in accessible terms, what the system does, what it uses, and how it may affect workers.

Le Quellenec

In practice, this cannot sit with one department. HR owns the use case; IT owns integration and security; the data protection officer (DPO) owns data protection and privacy; and legal owns the contracts. We recommend a small AI committee, a named owner for each tool, and an internal AI policy setting out authorized tools and uses.

The AI policy should also cover everyday generative AI. An HR team drafting job ads or appraisals with a chatbot is not high risk in itself. But pasting employee data into a public tool is a confidentiality and GDPR issue. Use approved tools only, and do not enter identifiable employee data unless the tool is contractually secured.

The vendor contract is central: instructions for use, information on training data and bias testing, accuracy metrics, access to logs, cooperation on incidents, and notice of updates. All of that is mandatory.

Updates matter because of the transitional rule. High-risk systems in use before December 2, 2027, only fall under the Act if their design is significantly changed afterward. Every major update could bring a tool into scope, so versions must be tracked.

Finally, plan for switching the tool off. If it proves noncompliant or discriminatory, the employer must be able to stop it, return to a human process, and keep the logs as long as needed for compliance and litigation before deleting them.

Neto Galvão

In relation to bias, bias is probably the biggest risk in HR. Where does it come from, and what can employers actually do about it?

Le Quellenec

That is not so easy to answer. It comes from several places. Historical bias: a tool trained on past hiring decisions reproduces past preferences. The well-known recruiting tool that penalized CVs mentioning women's organizations is an example.

Representation bias: some groups are underrepresented in the data.

Proxy bias: a postcode, a career gap, or a school name may be neutral on its face but stand in for origin, gender, or age.

Automation bias: a person validates whatever the tool suggests, turning human oversight into a rubber stamp.

There's also a source people rarely think about: the question itself. The French Lawyers' Guide on Ethics and AI, adopted in March this year, draws attention to it. The way you phrase a question shapes the answer, and generative tools tend to go along with its premise. Ask for the best candidate for a young, dynamic team, and you have introduced age discrimination before a single CV has been read. Ask whether a candidate is a good cultural fit, and you are asking the tool to reward similarity. HR prompts should be treated like a selection grid: neutral, written down, validated, and reviewed.

Neto Galvão

To correct bias, you sometimes need sensitive data, which is where the Omnibus helps. The new rules allow providers and deployers of all AI systems to process special categories of data to detect and correct bias, but only where strictly necessary and where synthetic or anonymized data will not do.

That requires pseudonymization, security measures, and deletion once the bias is corrected. It is a narrow door, not an open door.

Le Quellenec

On top of that, employers should test outcomes, not just accept vendors' assurance that the tool is fair. That means checking whether selection, promotion, or performance outcomes differ materially between groups, investigating the reason for any disparity, and documenting any corrective action.

Testing should also be repeated after significant updates or changes in the workforce or the use case. I was walking down the streets of Paris this weekend and saw an advertisement from a famous AI provider that I want to quote: "In HR, a 99% accurate AI model is a 100% fail rate."

Neto Galvão

Turning to AI literacy, which is the fourth step, the Omnibus softened the obligation. But how far, and what does it mean for employers?

Le Quellenec

The original text required companies to ensure a sufficient level of AI literacy among their staff. Now, they must take measures to support it. Lawyers will recognize the shift from an obligation of result to an obligation of means.

Neto Galvão

But an obligation of means is still an obligation, and it follows an accountability logic. The employer must be able to prove it took those measures. That means identifying who uses which tools, training them in proportion to the risk, and keeping records of who was trained, when, and on what.

For high-risk tools, the bar is higher. The AI Act still requires the people exercising oversight to have the necessary competence, training, and authority. Put simply, a recruiter who cannot explain why a candidate was ranked low is not supervising anything.

Le Quellenec

I would add that training should be role-based. A recruiter needs to understand ranking errors and discriminatory proxies. A manager needs to know when not to rely on an AI recommendation. Procurement needs to ask the right vendor questions. Senior management needs enough information to allocate responsibility and resources. One generic awareness session will rarely be enough.

Neto Galvão

Now, turning to recruitment in practice, let's take the case clients ask about most: recruitment itself. How should employers frame these tools in practice?

Le Quellenec

Everything depends on the purpose. If the tool stores and organizes CVs, there is no high-risk issue. If it helps choose between candidates, such as by scoring, ranking, or shortlisting, it is high risk. The purpose must be defined precisely, recorded, and respected in practice. That last part is the hard one.

Imagine recruiters pasting CVs into a general-purpose chatbot and asking it to rank them. The company has arguably put a general-purpose tool to a high-risk use and may be treated as the provider as a result. Hence the need to control uses and prompts.

On the GDPR side, a DPIA will, in practice, be required. Evaluating and scoring candidates, often at scale, meets the threshold for that assessment, and the French data protection authority's list specifically covers profiling for HR management.

Neto Galvão

Then there is the question we always get: Should we ask candidates for their consent? Generally, no. It would be a fragile basis. The European Data Protection Board's (EDPB) guidelines are clear that, given the imbalance of power in employment, consent will rarely be freely given. A candidate who refuses may fear losing the opportunity, and consent can be withdrawn at any time.

There are better bases, such as pre-contractual measures or legitimate interests with a documented balancing test. If the decision is fully automated, the GDPR rules on automated decision-making will apply. It is allowed only if necessary for the contract, authorized by law, or based on explicit consent. The person then has the right to human intervention, to express their views, and to contest the decision.

Before we close, what does a client really risk in relation to sanctions?

Le Quellenec

Up to 35 million euros or 7% of worldwide turnover for prohibited practices, such as the emotion recognition example. Up to 15 million euros or 3% for breaches of employer obligations, including transparency. For small and medium-sized enterprises, the lower of the two amounts applies.

Transparency has already been in force since August of this year. An HR chatbot answering employees' questions must make clear that it is AI, and AI-generated content must be marked.

Neto Galvão

All of this comes on top of GDPR fines and employment litigation. A discriminatory tool is first and foremost a discrimination case. Eric, if you had to sum it up?

Le Quellenec

Five words: map, govern, debias, train, frame. Map your HR AI tools, starting from the GDPR record. Govern them through rules, contracts, and the ability to switch them off. Tackle bias, including in prompts. Train your people and keep proof. Frame recruitment uses precisely, because that is where the risks spike first.

Neto Galvão

One last thing: the best compliance plans are built around the employment decision, not around the technology. Start with the decision, whether hiring, promotion, dismissal, task allocation, or performance assessment, and then identify where AI influences that process.

Make sure there is a lawful purpose, reliable data, meaningful human review, a route for the individual to challenge the outcome, and evidence that the process has actually worked.

To sum it up, the Omnibus has given employers more time, but it has not changed the destination. We have one year to do everything we have discussed in this podcast. Thank you very much for listening. 

© Jackson Lewis P.C. This material is provided for informational purposes only. It is not intended to constitute legal advice nor does it create a client-lawyer relationship between Jackson Lewis and any recipient. Recipients should consult with counsel before taking any actions based on the information contained within this material. This material may be considered attorney advertising in some jurisdictions. Prior results do not guarantee a similar outcome. 

Focused on employment and labor law since 1958, Jackson Lewis P.C.’s 1,100+ attorneys located in major cities nationwide consistently identify and respond to new ways workplace law intersects business. We help employers develop proactive strategies, strong policies and business-oriented solutions to cultivate high-functioning workforces that are engaged and stable, and share our clients’ goals to emphasize belonging and respect for the contributions of every employee. For more information, visit https://www.jacksonlewis.com.