AI in recruitment: screening CVs with proper safeguards

What the Spanish DPA said about using AI to analyse CVs

On 23 September 2026 the Spanish Data Protection Agency (AEPD) published a warning addressed to a company that was evaluating the deployment of an artificial intelligence tool for screening and evaluating job applications. The case file is EXP202600427 and the document is published as AI-00009-2026, signed by the Agency's president, Lorenzo Cotino Hueso.

It is worth starting with what the document is not, because it has been reported in several different ways.

What an Article 58(2)(a) warning actually is

It is not a fine. It does not declare that the company infringed the GDPR. Article 58(2)(a) allows a supervisory authority to issue a warning when intended processing operations are likely to infringe the Regulation. The decision itself dwells on that nuance: the warning belongs to a “precautionary” or “preventive” function, and the authority may issue it “not as a statement that those operations do in fact infringe the GDPR, but that there is a possibility (as opposed to certainty) that those operations may infringe the Regulation”.

The tool, moreover, had not yet been deployed in Spain. It was an initiative of the corporate group the company belongs to, developed centrally, and the Spanish company had confirmed that it would analyse the group's findings and adopt appropriate measures before any eventual go-live.

One detail in the file explains why this decision is more interesting than the headline. The company had already notified the workers' legal representatives of the deployment on 1 December 2025, under Article 64.4 of the Spanish Workers' Statute, stating that it was a support tool and that the final decision in the process is always human. In other words: it had done part of the work. The AEPD notes that fact and replies that the human intervention “must be effective”.

What the tool was intended to do

According to the decision, the system was intended for CV analysis, score assignment and the prioritisation of applications in personnel selection and internal mobility processes, and could “directly influence decisions relating to access to employment and professional promotion”. The company explained that its purpose was to improve the efficiency, consistency and objectivity of the initial evaluation, that the score reflected the degree of fit with the role, that the final screening or rejection decision rested with a person, that the system excluded the analysis of sensitive attributes and that it was subject to periodic audits.

The Agency does not dispute any of those statements. What it does is remind the company that none of them, on its own, closes the analysis.

The safeguards the decision sets out

The warning sets out four requirements. We summarise them in the document's own terms, adding nothing.

The decision closes by recalling that, if the necessary measures are not adopted, the company could commit an infringement leading to investigative or corrective action, including sanctions. That is the only point where the word sanction appears, and it appears as a conditional future scenario.

What the decision does not say

It does not say that using AI in recruitment is unlawful. On the contrary: it explicitly states that these tools “may offer opportunities to improve efficiency and consistency” and that, “properly designed and used”, they can improve the quality of certain processing operations. It does not say that every such tool requires a DPIA: it conditions that on the high risk of Article 35. It does not say that this particular case falls under Article 22: it says that must be assessed against how the process actually works. This decision is based on the GDPR and does not assess the system’s classification under the AI Act.

From here on, this is our proposed way of working, which is a different thing from the decision.

Not all AI in recruitment does the same thing

The first mistake would be to start with the tool. Before asking whether it complies with the AI Act, we should be able to explain what we want it to do. And “using AI in recruitment” describes very different things: drafting a job advert and automatically rejecting applications share the technology and little else.

This scale orders eight uses by their capacity to affect a specific person. It is not an AI Act classification, nor any authority's: it is our own explanatory tool for framing an internal conversation before turning to the law.

Levels 1 to 4 also require a review of purpose, data and actual influence on selection: a summary or comparison can shape a decision. Scoring or ranking people usually makes that influence more visible. At level 8, distinguish transcription and content analysis from inferring emotions from biometric data. The latter is prohibited in the workplace, except for medical or safety reasons; adding human oversight does not remove that prohibition.

The journey of an AI-assisted decision

An AI-assisted recruitment decision does not happen at a single point: it travels through seven stages, and a different discipline is involved at each one. This is the map we use to place responsibility before talking about tools.

The stage most often forgotten is the last one. A recruitment decision can be questioned months later, when nobody remembers which version of the system was live or with what configuration. If no trace remains, the organisation cannot demonstrate anything: neither that it did things properly nor that the decision had the reasons it claims.

What data the system ends up receiving

When people talk about “giving the CVs to the AI”, they are describing only the first column of this map. In practice, a screening system integrated into an applicant tracking system (ATS) sees considerably more, and returns things nobody gave it.

Provided data and inferred data are not the same thing

A candidate knows what they wrote in their CV. They do not know that a system has inferred a “seniority level” of 3 out of 5 from it, nor that this inference then travels to a spreadsheet. Both are personal data and both create obligations, but they behave differently: an inference cannot be verified by reading the original document, it usually lacks traceability and it is almost never held to the accuracy principle of Article 5(1)(d) with the same rigour as a declared field.

Four questions follow from that map, worth answering in writing before connecting anything: what the system genuinely needs for its purpose (minimisation), what else its output will be used for (purpose limitation), how a wrong inference gets corrected (accuracy) and how long each layer is kept (storage limitation). Scores tend to outlive the CVs they came from, and that is a retention problem nobody consciously decided.

GDPR: from articles to business decisions

This table does not reproduce the Regulation. It translates each applicable obligation into the question a management team has to be able to answer when someone asks.

Article 22, and why a score can already be the decision

There is a comfortable reading of Article 22 according to which it is enough for a person to sign at the end for the provision not to apply. That reading has a problem, and it is not theoretical.

In SCHUFA (Case C-634/21, 7 December 2023), the Court of Justice of the European Union held that the automated generation of a probability value itself constitutes an automated individual decision under Article 22(1) where that value determines in a decisive way whether a third party establishes, implements or terminates a contractual relationship with that person. The reasoning was built on credit scoring, but the mechanics are recognisable: a system produces a number, another actor receives it and almost always acts accordingly.

Applied to recruitment, the question is not whether there is a person at the end of the process. It is how much the score determines what that person does. If, out of 5,000 applications, only the 50 highest-scoring ones are reviewed, the decision about the other 4,950 was taken by the system, regardless of who signs the record.

There is a second relevant ruling. In Case C-203/22 (Dun & Bradstreet Austria, 27 February 2025), the Court clarified what “meaningful information about the logic involved” means: the controller must enable the person to understand the reasoning that led to the decision, and the supplier's trade secret does not work as an automatic exemption; it falls to the authority or the court to weigh the interests. For a company buying a tool, that has a very concrete practical consequence: if the supplier will not explain how it works, the transparency problem ends up being yours, not theirs.

When risk assessment stops being optional

Analysing risk is a general obligation flowing from Articles 24 and 32: it must always be done. The DPIA of Article 35 is something else: a formal procedure required where the processing is “likely to result in a high risk” to rights and freedoms. Not every recruitment AI requires one automatically, and it is worth not repeating that shortcut.

1. 01Are we processing personal data? In recruitment, always.
2. 02Is there evaluation, scoring or profiling?
3. 03Does the decision have significant effects on the person?
4. 04Do other criteria apply: scale, sensitive data, novel technology, people in an imbalanced position?
5. 05If the picture points to high risk: DPIA before the processing.

The criteria to cross-check are those of Article 35(3), the European Data Protection Board's guidelines on impact assessment and the list of processing operations the AEPD publishes for the Spanish market. A candidate is, by definition, in an imbalanced position relative to whoever decides whether they get a job, and a systematic evaluation with a score that governs access to a role brings several of those criteria together at once. That does not make the DPIA automatic, but it does mean the opposite conclusion has to be reasoned and written down.

Employment law: what must be disclosed, and to whom

This is the part that tends to be missing when the project is led by technology or compliance without HR in the room. There are two distinct obligations, with distinct recipients, and they should not be conflated.

Article 64.4.d of the Workers' Statute

The works council has the right to “be informed by the company of the parameters, rules and instructions on which the algorithms or artificial intelligence systems are based that affect decision-making which may have an impact on working conditions, access to and retention of employment, including profiling”.

Three points that matter. First: the provision expressly mentions access to employment, so it reaches recruitment processes and not only decisions about existing staff. Second: Article 64.4.d establishes a right to information and does not confer a veto. This does not exclude consultation or prior-report rights arising under other provisions, including Article 64(5), or the applicable collective agreement. Third: what has to be handed over are the parameters, rules and instructions, not a commercial description of the tool. In the AEPD case, the company had notified the deployment in December 2025, and the Agency still considered there were grounds to warn about the remaining safeguards.

Royal Decree 723/2026: who it reaches and who it does not

Royal Decree 723/2026 of 9 September, published in the Spanish Official Gazette on 15 September and entering into force on 5 October 2026, partially transposes Directive (EU) 2019/1152 on transparent and predictable working conditions and develops Article 8(5) of the Workers' Statute. Among the information the company must provide, Article 3(2)(k) lists “the existence of algorithmic or automated decision-making systems”, including “the guidelines, criteria and operating rules” where they are used to decide on the setting or modification of working conditions.

Precision matters here, and extrapolation does not help. The recipient of this obligation is the worker, within the information about their employment relationship. It does not regulate the recruitment stage and does not create an information right for someone who is still an external candidate. It does reach decisions about people already on the payroll, and that is where it meets the AEPD case: that tool was also intended for internal mobility. The same system may fall outside this royal decree when screening external applications and inside it when ranking internal promotions.

Collective agreements and bargaining

Some collective agreements and framework accords have begun to include clauses on the use of algorithms and artificial intelligence, with commitments on information, participation or prior assessment. It is worth reviewing the applicable agreement before designing the process, because it may add steps the law does not require. But the distinction should hold: a clause in a collective agreement binds those within its scope; a recommendation in a cross-industry accord guides bargaining and is not a legal rule. Mixing the two in an internal report is a reliable way to lose credibility with whoever has to approve the project.

AI Act: what applies today and what arrives in 2027

Regulation (EU) 2024/1689 applies in stages, and Regulation (EU) 2026/1744, the digital omnibus on AI, moved some dates in July 2026. Using a calendar from before that change leads to wrong conclusions in both directions: believing the high-risk regime is already upon you, or believing nothing applies yet.

What is already enforceable in September 2026

Three things, and the first matters most for recruitment.

The prohibition on inferring emotions. Article 5(1)(f) prohibits placing on the market, putting into service or using AI systems to infer emotions in the workplace and education, except for medical or safety reasons. It has applied since 2 February 2025. The definition in Article 3(39) and the Commission’s guidelines delimit this case to inference based on biometric data, such as voice characteristics or facial expressions. The workplace includes recruitment. Analysing those signals to infer enthusiasm in an interview falls within the prohibition; transcribing what was said or analysing text content alone does not automatically do so. That does not establish that those other uses are lawful: they must be assessed under the GDPR, the rest of the AI Act and employment and equality law.

AI literacy. Article 4 requires measures to support a sufficient level of AI literacy among the staff dealing with the operation of these systems. The omnibus softened the wording, from ensuring to supporting, but did not remove it. A recruitment team using a tool without understanding its limits is precisely the situation the provision has in mind.

Article 50 transparency. Since 2 August 2026 people must be told when they are interacting with an AI system, unless it is obvious. If the first filter in your process is a conversational assistant asking the candidate questions, this applies to you today.

Annex III, point 4: which employment systems are covered

Annex III lists the use cases that Article 6(2) treats as high risk. Its point 4, on employment, workers management and access to self-employment, covers two cases: systems intended for the recruitment or selection of natural persons, “in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates”; and systems intended to decide on terms of work-related relationships, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance.

A CV screening system with scoring sits squarely within the first limb. But claiming that “all HR AI is high risk” is still wrong, because of what follows.

The Article 6(3) exceptions and the profiling catch

Article 6(3) allows a system listed in Annex III not to be considered high risk where it does not pose a significant risk to health, safety or fundamental rights, and one of these four conditions is met: it is intended to perform a narrow procedural task; it improves the result of a previously completed human activity; it detects decision-making patterns or deviations without replacing or influencing the human assessment without proper review; or it performs a preparatory task for an assessment.

Now the decisive catch: those exceptions do not apply where the system performs profiling of natural persons. A system that scores applications by fit to a role is evaluating personal aspects to predict work performance, which is precisely the definition of profiling in Article 4(4) GDPR. In practice, that closes the exit route for most scoring-based screening tools. A field extractor turning a PDF into structured data can be argued under limb (a); an engine that ranks people, hardly.

It is worth adding that a provider relying on one of those exceptions must document its assessment and register it before placing the system on the market, under Article 6(4). If a salesperson tells you their product is not high risk, that documentation is exactly what you should ask for.

What changes on 2 December 2027

From that date, the high-risk systems of Annex III become subject to Chapter III: a risk management system (Art. 9), data governance and training-set quality (Art. 10), automatic event logging (Art. 12), instructions for use and transparency towards the deployer (Art. 13), human oversight built into the system itself (Art. 14) and accuracy, robustness and cybersecurity (Art. 15). For whoever buys and uses the tool, Article 26 adds its own duties: use it according to the instructions, assign oversight to people with the competence and authority to exercise it, control the relevance of input data, keep the logs and inform affected workers before putting it into service.

That is the list of duties fourteen months from now. It is also, under another name, much of what any organisation should already be able to explain about a system that decides who gets hired.

“A person decides” is not always enough

It is the sentence that appears in every product sheet and in nearly every internal report. It also appeared in the notice the company in the case file gave its workers' representatives. And it is probably where the distance is greatest between what a procedure says and what happens on a Tuesday afternoon with 300 applications pending.

What the evidence says about automation bias

That people tend to accept what an automated system proposes is not an intuition: it has been measured, and in recruitment settings.

Lacroux and Martin-Lacroux studied 694 professionals involved in screening applications and found that the relationship between what recruiters say about their trust in algorithmic recommendations and what they do with them is not the expected one: the consistency of the recommendation affected their decisions in unexpected ways, beyond their declared distrust of the algorithm (Frontiers in Psychology, 2022).

More uncomfortable still is the work of Kupfer and colleagues, who tested two ways of reducing automation bias in a personnel pre-selection task. Telling participants the system could make errors increased how intensively they verified information, but did not improve the objective quality of their decisions. And the condition emphasising their legal responsibility under the GDPR did not differ from the control group on any measure (Frontiers in Psychology, 2023). Put bluntly: reminding someone that they are responsible is not enough to make them review better.

On the regulatory side, Ben Green has documented how policies requiring human oversight can end up legitimising flawed systems and diluting accountability, by treating as settled a safeguard that in practice is not met (Computer Law & Security Review, 2022). That is not an argument for banning; it is an argument for not taking oversight on trust.

Seven conditions that hold human oversight up

Seven checkable conditions follow from that evidence and from the AEPD's requirement. They are not legal requirements listed in any law: they are our proposal for how to demonstrate what the law asks for.

1. Competence. The reviewer understands what the score measures, what it was validated against and which errors it makes most often.
2. Information. They can reach the full application, not just the output or the ranking position.
3. Time. The time allocated per application is compatible with reviewing it. If it is 40 seconds and 300 applications, the arithmetic answers the question.
4. Authority. They can depart from the output without asking a more senior level.
5. No penalty. A reasoned disagreement is not penalised or made impracticable by organisational barriers. Documenting the reasons for a decision remains appropriate.
6. Traceability. The review and its reason are recorded where appropriate, with a defined retention period.
7. Observability. Decision samples, detected errors and disagreements with the system are reviewed. The disagreement rate helps investigations; there is no minimum quota that demonstrates effective review.

Agreement with the system does not by itself demonstrate ineffective oversight, just as disagreement does not prove an adequate review. Check what information was examined, against which criteria and with what real ability to change the decision.

Bias: removing the “sex” variable does not close the problem

A supplier's most common defence is that their system “does not analyse sensitive attributes”. It even appears in the AEPD file. It is a necessary measure and an insufficient one, because bias rarely travels in the variable that bears its name.

Proxy variables

A proxy variable is an apparently neutral field that correlates with a protected characteristic. Postcode can correlate with origin or socioeconomic background. Career breaks correlate with caring responsibilities and therefore with gender. The university attended correlates with socioeconomic background and sometimes with origin. Years of experience correlate with age, almost mechanically. Register of language in a cover letter correlates with education and with first language. Willingness to travel correlates with family responsibilities.

None of those variables is discriminatory in itself. All of them are reasonable in some context. The problem appears when they combine, when the model learns from historical decisions that were already skewed, and when nobody measures the aggregate effect on specific groups. A system can see nobody's sex and still reproduce, with notable precision, the composition of a workforce built over twenty years of human decisions.

Where bias comes from

Distinguishing the origin helps identify which control applies, and that is the contribution of Suresh and Guttag's framework on sources of harm across a model's life cycle.

A fourth effect belongs to systems in production: feedback loops. If the system prioritises profiles resembling those already hired, those are the ones who get hired, and those hires feed the next retraining. Köchling and Wehner's systematic review of 36 studies on algorithmic discrimination in recruitment and people development documents this ground in more detail than fits here (Business Research, 2020).

Technical fairness and legal non-discrimination are different questions

Formal fairness metrics exist — demographic parity, equality of opportunity, calibration across groups — and they are useful. It is also established that several of them cannot be satisfied simultaneously except in trivial cases, so choosing a metric is a normative decision, not a technical one.

What matters for a management team is this: a fairness report with good numbers does not establish that a process is legally non-discriminatory, and a process can be problematic in equality terms without any metric flashing red. They are different planes with different logics. Raghavan and colleagues analysed what suppliers of algorithmic pre-employment assessments actually disclose about their development, validation and bias mitigation, and found a considerable gap between what is claimed and what is documented (ACM FAT*, 2020). That is a good reason to ask for the documentation rather than the claim.

Controls for using AI in recruitment

Each control addresses a risk in this use case. “Proposed measure” sets out our recommendation; “Duty or reference” identifies the legal duty or voluntary framework informing it. A GDPR reference does not mean that the Regulation prescribes that exact measure: its suitability must be justified against the risk and context.

If three have to be picked to start with, in our experience they are 06, 12 and 20. Logging, because without it nothing can be demonstrated afterwards. Change management, because that is where control is lost quietly. And exported data, because that is where the incidents nobody planned for turn up.

Which framework is for what

They get mixed up often, and that confusion has consequences: turning NIST into a European obligation or ISO 42001 into a legal requirement misleads whoever has to approve a budget.

• Binds today
• Binds later
• Voluntary

None replaces the others, and none of the voluntary ones is mandatory. What is true is that organisations already running an AI management system or holding an AI policy arrive at this conversation with half the work done: they have an inventory, owners, evaluation criteria and somewhere to record the decision.

NIST AI RMF applied to a recruitment process

Its four functions are a reasonable way to order the work, always remembering that it is a voluntary, non-European framework.

And a fourth, MANAGE: what happens when the risk stops being acceptable. It is the one that is almost never written down. It should cover who can suspend the system, on what criteria, and what happens to processes already under way when that occurs.

Twenty questions we should be able to ask the supplier

A commercial conversation about these tools usually revolves around efficiency and reduced time-to-hire. These twenty questions move it onto different ground. If any answer is “that is proprietary” or “we cannot detail it”, that is information too: note it down and decide whether that opacity is acceptable.

This block of questions is reused, condensed and with checkboxes, in the downloadable PDF. If you keep a third-party inventory, it fits well with our supplier assessment methodology and with the security questionnaire generator we publish as an open tool.

The ACBSEC × USOARIT six-step review

It is not an official standard and does not pretend to be. It is the order in which we approach this problem when we look at it together, and its main virtue is that the tool appears at step five, not step one.

1. NeedWhat problem we want to solve and how we will know we solved it. Often the real problem is that the role is poorly defined.Usoarit · Management
2. ProcessExactly where the AI will sit within the recruitment process, and what happens to an application depending on the output.Usoarit · HR
3. DataWhat goes in, what comes out, what is inferred, how long it is kept and what could be left out without losing the purpose.ACBSEC · DPO
4. PeopleWhich candidates and workers it can affect, what they are told, when, and who answers if they ask.Usoarit · ACBSEC
5. Technology and riskHow it works, which supplier is involved, what controls exist and what evidence remains.ACBSEC · IT
6. DecisionDeploy, change the scope, limit the use or drop it. With a date, an owner and a written reason.Management

Step six has four possible outcomes, and three of them are not “deploy”. A review process that can only end in yes is not a review process.

Are you evaluating AI for recruitment?

Before deploying it, we can review the whole process with you: what you want to automate, what data is involved, how it affects candidates and recruiters, which obligations apply and what technical and organisational controls you need.

People · process · labour relations × ACBSECAI · privacy · security · governance

Assess an AI recruitment projectSee what Usoarit does →

A worked example: 500 people, 5,000 applications

The following case is fictional. It does not correspond to any client or real organisation, and the decisions shown are those of this scenario, not universal obligations.

An industrial company of around 500 people receives roughly 5,000 applications a year for technical and plant roles. The recruitment team is three people. A supplier offers them a tool that “helps select better candidates”: it reads CVs, extracts experience, compares skills, scores and ranks. The demo is good and the time saving obvious.

The uncomfortable finding was the fourth. On paper there was human oversight in 100% of processes. In practice, applications below position 40 in each process were excluded without review. The annual total affected would have to be calculated from the number of processes, their applications and the reviews performed; it cannot be inferred from the 5,000 annual applications.

The fifth also created work: the part of the process with the greatest exposure risk was not in the supplier's platform, but in a spreadsheet of names, phone numbers and scores that three people forwarded to each other by email.

And the third opened the longest conversation. Validating a system against the company's own past hires sounds reasonable and is precisely the mechanism by which a model learns to reproduce the historical composition of the workforce. It does not prove discrimination; it does require looking.

What to review before using AI in recruitment

The summary table. It crosses the area, the question to answer, the risk if it goes unanswered, the control that reduces it, who should own it and where it comes from.

Why this needs two disciplines

A review done only from security and privacy produces a technically correct report about a process nobody has questioned: data that perhaps should not be there gets encrypted properly, and access is controlled to a score nobody knows how to read. A review done only from human resources produces a well-designed process on top of a black box: criteria, roles and review points get defined without knowing what the system actually does or what gets recorded.

People · process · labour relations

• Selection process design
• Evaluation criteria and roles
• Human intervention and reviewer competence
• Organisation and labour relations
• Communication and recruiter training
• Talent management and internal mobility

×

ACBSECAI · privacy · security · governance

• AI governance and risk assessment
• Privacy from the technical side
• Access, traceability and architecture
• Supplier assessment and contracts
• Controls and information security
• ISO/IEC 42001 and ISO/IEC 27001

Responsible and defensible use of AI in recruitment

The AEPD warning illustrates this well: for a planned deployment, it recalled the need to ensure and demonstrate data protection by design, risk assessment and effective human intervention. This is a preventive reminder, not a finding that the company lacked those safeguards. Preparing and checking that evidence requires coordination between both disciplines.

Usoarit and ACBSEC are two distinct brands collaborating on this kind of project. Neither is a subsidiary of the other and neither replaces the other: each brings what it knows how to do.

Frequently asked questions

How we prepared this guide

This guide is joint work by ACBSEC and Usoarit. The AI governance, technical privacy, supplier assessment, controls and traceability side comes from Antonio Cebreiro Bernárdez, a computer engineer and cybersecurity consultant. The selection process, evaluation criteria, organisation, labour relations and human intervention side comes from Victoria Melissa Toledo Luna, a chartered labour relations graduate and founding partner of Usoarit, specialising in labour relations, human resources and business management.

The documentary starting point is the original AEPD decision, read in full rather than through press coverage. Cross-referenced against it are the GDPR text, the AI Regulation as amended by the digital omnibus of July 2026, Spanish employment law, distinguishing current obligations from those taking effect on 5 October 2026, the European Commission's guidelines on prohibited practices, two CJEU judgments on automated decisions, the governance and security frameworks cited, and six peer-reviewed works on automation bias, human oversight and algorithmic discrimination in recruitment.

Neither ACBSEC nor Usoarit acts as a legal adviser through this material. The guide is informational and does not constitute individual legal advice. Whether each obligation applies depends on the specific system, its configuration and its context of use: what is offered here are points worth reviewing, controls to consider and questions an organisation should be able to answer.

Published and revised on 24 September 2026. Revision 1.1: legal effective dates, scope of emotion inference, human oversight and the practical scenario; download updated. When the AI Regulation's application dates change or relevant new rulings appear, we will update the guide and note it here.

Sources and documentation

Are you evaluating AI for recruitment?

If you are at step one — still deciding what you want the tool to do — that is the best moment to talk. We review the whole process with you: what you want to automate, what data is involved, how it affects candidates and recruiters, which obligations apply and what technical and organisational controls you need.

People · process · labour relations × ACBSECAI · privacy · security · governance

Review our caseDownload the PDF first