A guide for builders · As of November 2026

Building AI for recruitment? You are probably building a high-risk system.

Employers receive more and more applications, and CVs and cover letters look more and more alike because candidates use AI. The reflex: let AI sort the pile. That is exactly where the legal work starts. This guide lays out the four risks, what they cost and how to design it properly.

The four risks

Four risks for anyone building AI for recruitment

Anyone building AI for recruitment and selection runs into four risks. Risks 2, 3 and 4 have applied since 2018.

Risk 1

High-risk AI

AI Act, Annex III

When it hits you

Your system analyses or filters applications, or evaluates candidates (Annex III, point 4(a)). Scoring, ranking and matching all fall under that. If the system evaluates or predicts characteristics of the candidate, that is profiling and every exception falls away (Art. 6(3), final sentence).

The pitfall

You become the provider sooner than you think. Whoever gives a general-purpose language model a high-risk purpose becomes the provider of a high-risk system (Art. 25(1)(c)). Deliver it under your name and that is you. If your client puts it into service under its own name, it is your client (Art. 3(3)).

How to get it right

Let AI structure, not judge. From free text to fixed fields, without a score and without inferred traits. And before the first line of code, record who is the provider and who is the deployer.

How we do it: our AI chat turns answers into form fields and has no opinion about the candidate.

Since when

2 December 2027. Anyone relying on an exception documents that before go-live (Art. 6(4)).

High risk is not a label. It is a project.

All of this must be in place before a high-risk system may go live. Fourteen parts for the builder, and a separate list for the client.

For the builder (provider)

1. Risk management system for the full lifecycle

Art. 9

2. Data governance, including bias examination

Art. 10

3. Technical documentation

Art. 11, Annex IV

4. Automatic logging

Art. 12

5. Instructions for use for the client

Art. 13

6. Human oversight by design

Art. 14

7. Accuracy, robustness and cybersecurity

Art. 15

8. Quality management system

Art. 17

9. Conformity assessment

Art. 43

10. EU declaration of conformity

Art. 47

11. CE marking

Art. 48

12. Registration in the EU database

Art. 49

13. Post-market monitoring

Art. 72

14. Reporting serious incidents

Art. 73

For your client (deployer, Art. 26)
  • Use in line with the instructions for use
  • Oversight by competent, trained people
  • Keep logs for at least six months
  • Inform workers' representatives in advance
  • Inform candidates that high-risk AI is used
  • Explain a decision when the candidate asks (Art. 86)
Changing something substantial?

Then the conformity assessment starts again (Art. 43(4)). A new model or a different purpose can already qualify. High risk is not a one-off hurdle but an ongoing obligation.

The risk everyone forgets

Risk 2

Automated decision-making

GDPR, Article 22

All eyes are on the AI Act. That has been postponed. Article 22 GDPR has not: it has applied since 2018, even when no AI is involved.

When it hits you

A decision without meaningful human intervention, with legal effects or a similarly significant impact. Automatically rejecting a job applicant is the textbook example (recital 71 GDPR). Such a decision is prohibited unless one of three exceptions applies.

Since when

25 May 2018. No transition period, no postponement.

Why it bites now

The score is the decision

If the human follows the score in practice, the score itself is the automated decision (CJEU, SCHUFA, C-634/21). A recruiter who clicks approve does not change that.

Agents make it feasible

An agent that screens, rejects and reports back on its own is no longer a technical problem. Legally, it is exactly what Article 22 is about.

It applies without AI too

A knock-out form with fixed rules falls outside the AI Act, but not outside Article 22. Staying out of high risk does not solve this risk.

The three exceptions (Art. 22(2))

Necessary for the contract

Requires you to show that no less intrusive route would do. A heavy burden of justification.

A legal provision

None exists for recruitment and selection in the Netherlands.

Explicit consent

Must be freely given. Someone who says no and drops out of the procedure as a result never had a choice.

Always required, whichever exception applies

The right to human intervention, to express a point of view and to contest the decision (Art. 22(3)). And an explanation up front that a decision is automated, on what basis and with what consequences (Art. 13(2)(f)).

How to get it right

Choose deliberately. Either a human who can and may deviate, with all the data in front of them. Or an honest automated decision: hard, job-relevant criteria the candidate knows in advance, explicit consent and a human route without detriment for those who say no.

Our line: explicit consent, and whoever says no gets a human assessment.

Risk 3

Special category data

GDPR, Article 9

When it hits you

Health, origin, religion, trade union membership, biometrics. Processing is prohibited unless an exception applies. In recruitment this data enters in three ways.

Since when

25 May 2018.

Unrequested, through the CV

A photo, a period of illness, a union role. It is there, so your model processes it.

Inferred by the model

The Court reads Article 9 broadly: data that indirectly reveal sensitive information are covered too (C-184/20). A model that infers health or personality from text or voice creates special category data itself.

Through a test

Whether psychometric data are health data is not legally settled. We treat them as such.

The pitfall

The link with risk 2. An automated decision based on special category data is only possible with explicit consent or a substantial public interest (Art. 22(4)). The contract route is then closed.

How to get it right

Do not ask for it. Filter it out before the model sees it. Let the model infer nothing. And never make special category data a knock-out criterion.

Risk 4

The regular GDPR homework

GDPR, general

When it hits you

Always. Every tool that processes candidate data falls under the GDPR. AI adds one thing: the data often goes straight to a model vendor's API.

  • A valid legal basis for every processing activity (Art. 6)
  • A DPIA, completed before go-live (Art. 35)
  • Clear information for candidates (Art. 13 and 14)
  • Data processing agreements with every vendor in the chain (Art. 28)
  • No transfers outside the EEA without safeguards (Chapter V)
  • Process only what is needed, and keep it no longer than needed (Art. 5)

The pitfall

Forgetting your client's homework. Your client is the controller and must be able to demonstrate all of this. In the Netherlands, the works council also has a right of consent on arrangements about hiring policy (Art. 27 Works Councils Act).

How to get it right

Send only what is needed to the model, process within the EU and deliver the substantiation along with the system: the texts for candidates, the input for the DPIA and the explanation for the works council and the privacy officer.

Since when

25 May 2018.

And what is never allowed

Prohibited AI: emotion recognition in job applications

A video interview that weighs tone or facial expression is emotion recognition in the workplace. That has been prohibited since 2 February 2025 (Art. 5(1)(f) AI Act), and the Commission counts job applications as part of that. The fine: up to €35m or 7%. And since 2 August 2026, any AI that talks to people must say it is AI (Art. 50).

When it goes wrong

The fine per risk

Risk 1

High-risk AI

€15m or 3%

Art. 99(4) AI Act

Risk 2

Automated decision-making

€20m or 4%

Art. 83(5) GDPR

Risk 3

Special category data

€20m or 4%

Art. 83(5) GDPR

Risk 4

The regular GDPR homework

€20m or 4%

Art. 83(5) GDPR. For a missing DPIA or data processing agreement: €10m or 2%

Maximum fines, whichever is higher, as a percentage of worldwide annual turnover. Under the AI Act, SMEs and small mid-caps pay the lower of the two. The AI Act and the GDPR apply side by side (Art. 2(7) AI Act): one tool can be fined on several risks at once.

And the fine is not the only thing.

An authority can have a system withdrawn from the market (Art. 79 AI Act) or stop the processing (Art. 58 GDPR). Rejected candidates can claim damages (Art. 82 GDPR). And your client will then look at whoever built it.

Run the check

Ten questions, and what to do when the answer is wrong

Go through them for every system you build. For each answer shown below, you have work to do before you go live.

1

Does the model assess, score or rank candidates?

If yesRisk 1

Then you are building high-risk AI. Take the judgement out of the model, or plan the fourteen parts before 2 December 2027.

2

Does the model infer traits the candidate did not provide personally?

If yesRisk 1 and 3

That is profiling: no exception to high risk, and possibly special category data. Limit the model to what the candidate provides.

3

Does the candidate see what the AI made of the answers, and can they change it?

If noRisk 1

Then the AI steers the outcome out of the candidate's sight. Build in a step where the candidate checks, edits and submits personally.

4

Is a decision taken without a human assessing the substance?

If yesRisk 2

Then Article 22 GDPR applies, even when fixed rules decide. Arrange an exception, the safeguards and an explanation up front.

5

Can the human in the process genuinely deviate, with all the data and enough time?

If noRisk 2

Then it is an automated decision after all. Give the reviewer authority, all the data and time, or set it up in line with Article 22.

6

Is it recorded which exception an automatic rejection rests on, and is there a human route without detriment?

If noRisk 2

Then the rejection is prohibited. Choose the exception, record it and build the human route before you reject automatically.

7

Can special category data reach the model or the decision rules?

If yesRisk 3

Filter CVs before the model sees them, do not ask for such data and never use it as a knock-out criterion.

8

Is it on paper who is the provider and who is the deployer?

If noRisk 1

Then you may be carrying the provider obligations without knowing it. Record the division of roles in the contract, before the build.

9

Is the DPIA complete and is the works council involved?

If noRisk 4

Do not go live. Complete the DPIA first and put the deployment to the works council.

10

Does the system say it is AI, and does emotion recognition stay out?

If noProhibited AI

This already applies today. Add the AI notice and remove any analysis of tone, voice or face.

The pattern that works

AI prepares. The candidate confirms. Fixed rules decide.

Three steps, and in none of them does a model judge a human.

Step 1

AI prepares

The AI holds the conversation and turns free text into structured fields. A narrow, preparatory task: exactly what the exception in Article 6(3)(a) and (d) and recital 53 describes.

AI, not high risk

Step 2

The candidate confirms

The completed fields go back to the candidate, who checks, edits and submits them personally. Everything after that runs on data the candidate has confirmed.

Human

Step 3

Fixed rules decide

If-then rules that the employer sets in advance test the hard requirements. Same input, same outcome. Rules defined entirely by humans are not an AI system (Art. 3(1), recital 12). Article 22 GDPR continues to apply.

No AI

Where the pattern breaks

The Commission reads the exception narrowly (draft guidelines of 19 May 2026). The split only holds if the AI genuinely assesses nothing. A summary with a verdict, a hidden score or a model that interprets answers out of the candidate's sight, and you are back at risk 1. Document your assessment before you go live (Art. 6(4)) and register the system (Art. 49(2)). And risk 2 remains: fixed rules take an automated decision too.

How we built it

Selection Lab works according to this pattern. Our AI chat prefills a knock-out form that the candidate checks and submits personally. Our own assessments score with fixed keys and norm groups, without AI. That way you select on what someone can do and who someone is, not on a CV that looks like all the others.

Read the reasoning behind the AI chatRead how it works for assessments

What applies when

The GDPR is not waiting for anything. The AI Act arrives in steps.

25 May 2018

GDPR, including Article 22 and the DPIA duty

2 Feb 2025

Prohibited AI practices and AI literacy (Art. 4 and 5)

2 Aug 2025

AI Act penalty regime (Art. 99)

2 Aug 2026

Transparency duty (Art. 50)

2 Dec 2027

High-risk rules for recruitment and selection (Annex III)

Prefer to build on something that already exists?

Our AI chat and assessments are the building block under your solution. We will walk through your design with you in half an hour.

Book a call

Sources. Regulation (EU) 2024/1689 (AI Act), as amended by Regulation (EU) 2026/1744: Art. 2(7), 3(1), 3(3), 5(1)(f), 6(3), 6(4), 9 to 15, 17, 25, 26, 43, 47, 48, 49, 50, 72, 73, 79, 86 and 99, recitals 12 and 53, Annex III point 4(a). Regulation (EU) 2016/679 (GDPR): Art. 5, 6, 9, 13, 14, 22, 28, 35, 58, 82 and 83, Chapter V, recital 71. Court of Justice of the EU, 7 December 2023, C-634/21 (SCHUFA) and 1 August 2022, C-184/20. Dutch Works Councils Act, Art. 27. European Commission, guidelines on prohibited AI practices (4 February 2025) and draft guidelines on the classification of high-risk AI systems (19 May 2026).

This guide is general information as of November 2026 and not legal advice for your specific situation.