Before an AI tool goes live in recruiting in Germany, Austria or Switzerland, four checks are needed, and they run in parallel: works council co-determination (§ 87(1) no. 6, § 90, § 95(2a) BetrVG), data protection (GDPR Art. 6, 13/14, 22 and 35), classification under the EU AI Act (Annex III no. 4, deployer duties from 2 December 2027) and non-discrimination under the AGG. None of them replaces another, and no vendor can take the responsibility off you. This page puts them in order. It is general information, not legal advice.
| Track | Legal basis | Triggered when | What you need beforehand |
|---|---|---|---|
| Works council | § 87(1) no. 6 BetrVG, § 90, § 95(2a), § 80(3) | Tool can monitor behaviour or performance; AI shapes selection guidelines; AI is being introduced | Function map, system profile, draft works agreement |
| Data protection | GDPR Art. 6, 13/14, 22, 28, 35; § 26 BDSG | Always, once applicant data is processed | Legal basis, notices, processing agreement, DPIA |
| AI Act | Reg. (EU) 2024/1689, Annex III no. 4, Art. 26 | Tool selects, filters, evaluates or ranks candidates | Classification per function, vendor documents, oversight and logging concept |
| Equal treatment | AGG §§ 1, 7, 15, 22 | Every selection decision | Documented role-based criteria; spot checks |
Does the works council have to agree? Co-determination under the BetrVG
Not every AI tool automatically needs a works agreement, but almost every recruiting platform triggers at least one participation right. What counts is function, data and effect, not the label "AI".
- § 87(1) no. 6 BetrVG: co-determination on the introduction and use of technical systems designed to monitor employee behaviour or performance. Recruiting software logs who reviewed which profile or changed which score, which makes recruiters and hiring managers measurable even if nobody intends it. Courts treat objective suitability for monitoring as enough.
- § 95(1) and (2a) BetrVG: guidelines on personnel selection in hiring need the works council's consent. Since the 2021 reform this expressly applies when AI is used in setting up the guidelines. An AI ranking with fixed criteria and weights is in essence such a guideline.
- § 90(1) no. 3 BetrVG: the employer must inform and consult the works council in good time on planned work procedures, including the use of AI.
- § 80(3) BetrVG: if the works council must assess the introduction or use of AI, calling in an expert is deemed necessary. Plan time and budget.
- § 99 BetrVG: participation in individual hires remains. Approving the tool does not settle it.
A pilot with real applicants is not a legal sandbox. A two-level model works in practice: a framework works agreement plus a system profile per tool. Templates and arguments are in Works council and AI recruiting in Germany; the short answer to the approval question is in Does the works council have to approve AI in recruiting?. In Austria, § 96(1) no. 3 and § 96a ArbVG cover control measures and staff assessment systems; in Switzerland the Data Protection Act and Art. 328b of the Code of Obligations apply. Check your own set-up there.
What does the GDPR require before use?
Legal basis and information
Applicant data counts as employee data (§ 26(8) sentence 2 BDSG). Whether § 26(1) BDSG holds up in full after the CJEU ruling of 30 March 2023 (C-34/21) is disputed, so document Art. 6(1)(b) and (f) GDPR as well. A works agreement creates a framework, but it does not make unnecessary processing lawful.
Applicants must be informed under Art. 13. In active sourcing you do not collect the data from the person themselves, so Art. 14 applies: the information is due within one month at the latest, and no later than the first message if you contact the person (Art. 14(3)). See Active sourcing and GDPR.
Art. 22 GDPR: no purely automated rejection
People have the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects (Art. 22(1)). An automatic rejection falls under it. The exceptions (contract, law, explicit consent, Art. 22(2)) require safeguards: human intervention, the chance to state one's view and to contest (para. 3). The CJEU's SCHUFA judgment of 7 December 2023 (C-634/21) matters here: creating a score can itself be an automated decision where the recipient relies heavily on it. The case concerned credit scores. It has not been decided for applicant scores, but the reasoning is close. A score that recruiters merely wave through does not amount to human review.
The design rule that follows: a person with information, time and authority decides every rejection. The AI sorts or recommends; it does not reject.
Art. 35 GDPR: data protection impact assessment
A DPIA is needed where processing is likely to result in a high risk. Art. 35(3)(a) names systematic and extensive evaluation of personal aspects based on automated processing, including profiling. The German data protection authorities' conference lists AI used to evaluate personal aspects on its mandatory DPIA list. For screening, ranking and matching, assume a DPIA and use the vendor's instructions as input. You also need a concrete processing agreement under Art. 28 with a sub-processor list and, for third-country access, a transfer mechanism under Chapter V. EU hosting alone does not answer that.
AI Act: what applies to recruiting, and from when?
Recruiting AI is listed in Annex III no. 4(a) of the EU AI Act: systems for recruitment or selection, in particular to place targeted job ads, to analyse and filter applications and to evaluate candidates. The Commission's employment examples name job matching, ranking and sourcing on online platforms. Pure interview scheduling or employer marketing mostly falls out. Systems that profile candidates always stay high-risk under Art. 6(3).
| Rule | Content | Applies |
|---|---|---|
| Art. 5(1)(f) | Ban on emotion recognition in the workplace (exceptions for medical and safety reasons) | since 2 February 2025 |
| Art. 50 | Transparency: people must be told they are interacting with an AI (for example an AI voice interview) | from 2 August 2026 |
| Annex III no. 4, Art. 26 | High-risk duties for deployers: use as instructed, human oversight with competence and authority, logs for at least six months, informing workers' representatives and affected persons | 2 December 2027 (Reg. (EU) 2026/1744, in force since 27 July 2026) |
Many pages still give 2 August 2026 for high-risk systems. That is out of date. Despite the later date, Art. 22 GDPR, the BetrVG and Art. 50 AI Act apply already. Building the works council process, function map and documentation now means you lose no time in 2027. The duties in detail: EU AI Act recruiting: which obligations actually apply.
Equal treatment: why the AGG and bias checks belong here
The employer is liable for discriminatory selection results even when an algorithm produces them (§ 15 AGG). Under § 22 AGG, indications are enough and the burden of proof then shifts to the employer. So fix role-based criteria before use, ask for reasons per requirement and check results by sampling for unevenly distributed errors.
The order before rollout: 8 steps
- Narrow the use case and functions: search, outreach, ranking, interview. Each function gets its own classification.
- AI Act classification per function: high-risk or not, with reasons.
- Request vendor documents: processing agreement, sub-processors, instructions for use (Art. 13 AI Act), logging concept, test results. A verbal "compliant" does not count.
- Fix legal basis and notices: Art. 6, Art. 13/14, notice texts, and the AI disclosure for voice interviews.
- Carry out the DPIA and feed in the vendor's input.
- Involve the works council early: present the function map, plan for an expert, negotiate the agreement.
- Write down the decision boundary: who decides on rejections, with what information, in what time.
- Pilot in a limited way, then evaluate: duration, users, stop criteria, log of overrides.
What evidence should a vendor provide?
- A concrete processing agreement with sub-processor list and hosting location, including support access.
- Documentation of what the system evaluates and what it does not, with reasons per requirement rather than an unexplained overall score.
- Logging and retention that let you exercise oversight (at least six months under Art. 26 AI Act).
- Settings for approval before sending and for human decisions.
- A statement on whether your data is used for shared models.
- Support for the DPIA and for informing the works council.
Measure every vendor against this list, ourselves included, and get the documents in writing. Atlas is hosted in the EU; whether your specific use is permitted is still decided by your own review, and a "GDPR-compliant" label does not replace it. How an AI recruiter works day to day and where approval steps sit is in AI Recruiter 2026; the search itself is shown in Active Sourcing with Atlas.
Which contract clauses belong in an AI recruiting vendor agreement?
An AI recruiting contract should cover audit rights, bias testing, data handling and change notice, because the deployer carries duties the vendor cannot take over. Use this list in procurement:
| Clause | Legal anchor | What to ask for |
|---|---|---|
| Processing agreement and audit rights | Art. 28(3) GDPR: instructions, sub-processors, inspections and audits | Right to audit or to receive audit reports; sub-processor list with notice of changes |
| Instructions for use and documentation | Art. 13 and 26(1) AI Act: the provider supplies instructions, the deployer follows them | Technical documentation and intended purpose in writing; no use outside it |
| Logs and oversight | Art. 26(6) AI Act: keep logs at least six months | Log export without extra fees; named human review step |
| Bias and adverse-impact testing | AGG, Art. 10 AI Act (data governance) | Test method and results shared on request, repeated after material model changes |
| Change notice | Art. 26 AI Act, works council rights under BetrVG § 90 | Notice before scoring logic or model changes, so the works council can be informed |
| Deletion and exit | Art. 28(3)(g) GDPR: delete or return at the end | Deletion within a set period, written confirmation, no training on your data |
| Liability and incidents | Art. 33 GDPR: breach notification | Incident notice within hours, liability for regulatory findings tied to the tool |
This list is orientation, not legal advice; have your data protection officer and legal counsel review the final text. The order of the checks before rollout is in the eight steps above.
Frequently asked questions (FAQ)
Can an AI tool reject applicants automatically?
As a rule, no. Art. 22 GDPR protects against decisions based solely on automated processing with significant effect. The exceptions are narrow and require human intervention and a way to contest. Let the AI pre-sort and a person decide.
Is AI active sourcing high-risk under the AI Act?
Sourcing that evaluates, filters or ranks candidates falls under Annex III no. 4(a). A pure data search without evaluation needs a more nuanced view. Check each function separately and document the reasoning.
When do the high-risk duties for recruiting apply?
From 2 December 2027 under Regulation (EU) 2026/1744 (Digital Omnibus). The Art. 5 prohibitions apply since 2 February 2025, the Art. 50 transparency duty from 2 August 2026.
Does a pilot need the works council's consent?
That depends on the functions, not on the word "pilot". With real applicants or employee data the same co-determination triggers can apply as in regular operation. Agree scope, data, duration and stop criteria beforehand.
Is EU hosting enough for the GDPR?
No. EU hosting is a good signal, but legal basis, notices, processing agreement, DPIA and third-country access by support or sub-processors are separate questions.
Do I need a DPIA for AI screening?
For screening, ranking and matching, usually yes (Art. 35(3)(a) GDPR and the German authorities' mandatory list). Clarify it with your data protection officer.
As of 5 November 2026. General information, not legal advice. Have the specific use reviewed by your data protection officer and employment counsel.
