AI Recruiting AGG/BAG-Compliant: 8 Safeguards

Main article: Algorithmic discrimination in recruiting — the legal position in 2026 — the full pillar article on the topic.

Two areas of law, two review paths

AI-supported personnel selection is covered by two independent sets of rules: anti-discrimination law and the AI Act (Regulation (EU) 2024/1689). A system that fully meets the deployer obligations under Art. 26 AI Act is therefore not immune to a claim under § 15 AGG — and conversely, clean AGG documentation does not replace any of the AI Act obligations.

Anti-discrimination law operates in relation to the individual applicant and is activated retrospectively in the event of a dispute: in Germany, § 7(1) AGG prohibits disadvantage on a ground listed in § 1 AGG — including where the ground is merely assumed; in Austria, the Equal Treatment Act takes on this role. The AI Act, by contrast, attaches to the role of deployer, regardless of whether discrimination has occurred, and applies identically in both countries as a directly applicable EU regulation; under Art. 70(1) AI Act each Member State designates at least one market surveillance authority, one of them as single point of contact (Art. 70(2)). In Germany this has been the Federal Network Agency (Bundesnetzagentur) since 29 July 2026 (§ 2(1) KI-MIG), in so far as sectoral authorities are not responsible; as at the date of this article (17 August 2026), Austria has not designated a market surveillance authority under Art. 70 AI Act — the RTR AI service point is an information and advisory body without supervisory powers.

When recruiting AI is high-risk AI — and who the deployer is

Recruiting AI is classified as a high-risk AI system under Art. 6(2) AI Act in conjunction with Annex III no. 4; Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027 (Art. 113(3)(c)(i) AI Act as amended by Regulation (EU) 2026/1744, in force since 27 July 2026). No. 4(a) covers AI systems 'intended to be used 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'; (b) covers systems for decisions on terms of work-related relationships, promotions and terminations as well as for task allocation and performance evaluation.

The prior question is whether there is an AI system at all: Art. 3 no. 1 AI Act defines it as a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment and that infers, from the input it receives, how to generate outputs such as predictions, recommendations or decisions. The exception in Art. 6(3) AI Act applies only where there is no significant risk and is excluded as soon as the system performs profiling of natural persons; where individual applicants are scored or ranked, this route is therefore regularly closed.

The deployer is the employer using the system — regardless of whether it trains the model itself or obtains it from a provider. The provider supplies documentation and instructions for use; the deployer obligations under Art. 26 AI Act remain with the employer — they apply to Annex III systems from 2 December 2027 (Art. 113(3)(c)(i) AI Act as amended by Regulation (EU) 2026/1744, in force since 27 July 2026).

On the Digital Omnibus: the AI part was adopted as Regulation (EU) 2026/1744 of 8 July 2026 and has been in force since 27 July 2026; the Annex III dates therefore fall on 2 December 2027 and the Annex I dates on 2 August 2028 (Art. 113(3)(c) AI Act as amended).

The eight safeguards at a glance

The 'Cycle' column contains practice recommendations only; statutory deadlines are set out in the 'Trigger / deadline' column and are identified as such there.

1. Bias test for each AGG characteristic

Statistical analysis of age, gender, ethnicity. Important: § 22 AGG sets no percentage threshold — it is sufficient that there are 'indications giving rise to a presumption of disadvantage on a ground listed in § 1'; the employer then bears the burden of proving that no infringement occurred. The 5% threshold is therefore not a judicially established criterion for prima facie evidence, but a statistical convention in practice (significance level) for your own bias analysis. Practice recommendation: repeat quarterly (not a statutory deadline).

The scope of testing follows the statute. § 1 AGG names race or ethnic origin, gender, religion or belief, disability, age and sexual identity. In Austria the catalogue is drawn differently: § 3 GlBG covers gender, § 17 GlBG ethnic affiliation, religion or belief, age and sexual orientation — disability is not a GlBG ground and is dealt with under §§ 7a et seq. BEinstG. A German test matrix transferred unchanged assesses an Austrian site against the wrong provision.

The dangerous category is indirect disadvantage under § 3(2) AGG: apparently neutral criteria that are liable to put persons at a particular disadvantage. That is exactly how training-data bias from historical hiring decisions, proxy variables such as postcode, club membership or an unbroken employment history, and constructs such as 'cultural fit' or 'digital native', which in practice operate as age filters, take effect. A model can formally ignore every protected characteristic and still screen people out systematically. Do not confuse the issues: the 5% threshold in Art. 10(1)(a) of Directive (EU) 2023/970 concerns the joint pay assessment by employers subject to reporting obligations, not the bias test in the selection procedure — and for private employers it takes effect only through the national transposing law: the transposition deadline (7 June 2026) has passed, and as at 17 August 2026 no German transposing act was in place (BT-Drs. 21/7026); in Austria, no federal transposing act was in place as at the same date.

2. Anonymised initial screening

Hide personal details (name, photo, date of birth, nationality where applicable) during the initial screening. Reactivate them only at the final selection stage.

Why this holds up. The safeguard removes from the model the direct points of reference for a direct disadvantage under § 3(1) AGG and at the same time implements data minimisation under Art. 5(1)(c) GDPR: data not required for the technical pre-selection are not processed at this stage. Some of the information additionally falls under Art. 9(1) GDPR, which in principle prohibits the processing of data on racial and ethnic origin, religious or philosophical beliefs, trade union membership and health data (exceptions under Art. 9(2)).

Limits of the safeguard. Free text in cover letters and CVs frequently conveys such information unintentionally — masking structured fields alone falls short. Anonymisation ends at the interview at the latest; from that point, a structured interview, a score sheet and four-eyes review carry the evidential burden. Practice recommendation: anchor anonymisation as a protective measure in the data protection impact assessment.

3. Human final decision

The AI only makes a recommendation — a human takes the final decision. Art. 22(3) GDPR plus Art. 14 EU AI Act human oversight.

What the AI Act specifically requires. Under Art. 26(2) in conjunction with Art. 14 AI Act, the deployer must ensure effective human oversight by persons trained for that purpose — the responsibility must be named, staffed and qualified. Added to this are Art. 26(1) (operation in accordance with the provider's instructions) and Art. 26(5) in conjunction with Art. 13 (monitoring of operation on the basis of the instructions for use). The training requirement has a second root: § 12(1) to (3) AGG obliges the employer to take the necessary and preventive protective measures, with appropriate training of staff expressly counting as compliance; in addition there is Art. 4 AI Act in the version applicable since 27 July 2026 (Regulation (EU) 2026/1744): providers and deployers must take measures to support the development of AI literacy among their staff — the provision expressly does not require a particular level of competence. Art. 4 AI Act does not prescribe a certificate; practice recommendation: document the measures taken internally (proof of training).

Why this decides the case in court. The mere fact of an algorithmic rejection without human review can in itself be treated as an indication within the meaning of § 22 AGG; once the threshold is crossed, the employer must prove that there was no breach of the prohibition of disadvantage. Practice recommendation: document the review for each application with a time stamp, the reviewing person and reasons that remain comprehensible without the AI score.

4. Transparency for applicants

Privacy notice: 'We use AI-supported pre-selection.' Rights in the case of automated decisions: 'You can obtain human intervention, express your point of view and contest the decision.' (Art. 22(3) GDPR)

Information about the logic is the hard part. On 27 February 2025, in Case C-203/22 (Dun & Bradstreet Austria), the Court of Justice gave substance to the right of access under Art. 15(1)(h) GDPR: in the case of automated individual decisions within the meaning of Art. 22 GDPR, data subjects are entitled to meaningful information about the logic involved — such that the decision can be understood and reviewed. A blanket reference to trade secrets is not sufficient; where there is a conflict, the information must be disclosed to a competent authority or to the court, which weighs the interests. Vendor contracts without audit and information rights make such disclosure practically impossible.

Two information duties that are often conflated. Art. 26(7) AI Act obliges the deployer to inform workers' representatives and the affected workers before putting a high-risk AI system into service in the workplace; Art. 26(11) AI Act, by contrast, concerns information to natural persons who are subject to a decision of the Annex III system. There is a hard limit above this: emotion recognition in the workplace and in education institutions is prohibited under Art. 5(1)(f) AI Act — no amount of transparency helps there. Where a system interacts directly with applicants, Art. 50(1) AI Act applies in addition, obliging the provider to ensure that information about the AI interaction is given.

5. Duty to give reasons

For each applicant: written reasons for the selection or rejection (practice recommendation for the burden of proof under § 22 AGG). The AI score alone is NOT sufficient.

Why the score is not enough. Where indications are proven, § 22 AGG requires the employer to prove that there was no breach of the prohibitions of disadvantage. That proof is furnished through the documented selection criteria and their application to the individual case; a numerical value without any link to criteria shifts the burden of proof into a model that the employer would have to explain in court. In addition, Art. 26(6) in conjunction with Art. 12 AI Act requires the automatically generated logs to be kept for at least 6 months: the log shows what the system did, the reasons show why the human decided as they did.

Germany and Austria: different time windows. In Germany, a claim under § 15(1) or (2) AGG must be asserted in writing within two months; for an application, the period starts when the rejection is received (§ 15(4) AGG). In Austria, claims relating to the refusal to establish an employment relationship must be brought before the courts within six months of the rejection (§ 15 GlBG for Part I, § 29 GlBG for Part II). Anyone operating both sites should align the availability of documents with the longer window.

6. FRIA — mandatory only for public bodies, otherwise best practice (obligation for Annex III systems from 2 December 2027; Regulation (EU) 2026/1744, in force since 27 July 2026)

A common misunderstanding: Annex III no. 4 EU AI Act (employment) triggers no general FRIA obligation. Under Art. 27(1) AI Act, the fundamental rights impact assessment must be carried out only by deployers 'that are bodies governed by public law, or are private entities providing public services', as well as deployers of systems under Annex III no. 5 points (b) and (c) (creditworthiness, life and health insurance). A private-sector employer using recruiting AI is therefore not subject to the FRIA obligation.

What applies to that employer instead are the deployer obligations under Art. 26 AI Act — in particular paragraph 7 (information to workers' representatives and the affected workers before putting the system into service in the workplace) and paragraph 9 (use of the provider information under Art. 13 for the data protection impact assessment under Art. 35 GDPR). Public employers and private entities with a public service remit will carry out the FRIA for Annex III systems from 2 December 2027 (Art. 113(3)(c)(i) AI Act as amended by Regulation (EU) 2026/1744, in force since 27 July 2026); for everyone else, the 7-step template remains voluntary best practice that usefully complements the DPIA.

The DPIA is the obligation that actually bites. Where the FRIA obligation exists, it must be fulfilled before the system is first used (Art. 27(2) AI Act) — classification as an addressee of the obligation is therefore the first check when purchasing a system. Independently of that, Art. 35(1) GDPR requires a data protection impact assessment in advance where a type of processing, in particular using new technologies, is likely to result in a high risk; Art. 35(3) expressly names systematic and extensive evaluation of personal aspects (profiling) producing legal effects. Applicant scoring regularly falls within this area; whether the threshold is reached must be assessed for each processing operation.

A knock-on effect in Germany, but not in Austria. Where processing is subject to a DPIA under Art. 35 GDPR, a data protection officer must be designated in Germany irrespective of the number of employees (§ 38(1) BDSG, alongside the 20-person threshold). Austria does not provide for any numerical threshold going beyond Art. 37 GDPR; the designation obligation there follows directly from Art. 37 GDPR, supplemented by § 5 DSG.

7. Complaints mechanism

A clearly communicated complaints route for applicants. Response within 1 month (Art. 12 GDPR).

The one-month period, precisely. Art. 12(3) GDPR requires information on the action taken on a request under Art. 15 to 22 without undue delay and in any event within one month of receipt. The period may be extended by two further months where necessary, taking into account the complexity and number of the requests — but the person must be informed of the extension within the first month, together with the reasons. The extension is therefore itself subject to a deadline, not a silent option.

The second channel. Under § 13(1) AGG, employees have the right to complain to the competent bodies of the establishment, the undertaking or the agency if they feel disadvantaged on a ground listed in § 1 AGG; the complaint must be examined and the outcome communicated. The statute and the competent body must be made known within the establishment (§ 12(5) AGG). The prohibition of victimisation applies alongside this: § 16(1) AGG also protects supporting persons and witnesses, in Austria § 13 GlBG (Part I) and § 27 GlBG (Part II). Which body outside the company is competent differs by country — see the comparison below.

8. Retraining and update cycle

Where bias is identified: retrain or replace the AI model. The update frequency is documented for each provider.

Why updates are a regulatory event. Art. 26(1) AI Act requires operation in accordance with the provider's instructions, and Art. 26(5) in conjunction with Art. 13 requires monitoring of operation on the basis of the instructions for use. A model update potentially changes the very basis on which the last bias analysis and the last oversight documentation rested. The cycle closes here: update → renewed bias test (safeguard 1) → updated oversight and reasoning documentation (safeguards 3 and 5).

Secure the provider interface contractually. Where application data are processed at the provider, the data processing agreement under Art. 28 GDPR is the basis: the controller may use only processors providing sufficient guarantees to implement appropriate technical and organisational measures, and the contract must set out the subject matter, duration, nature and purpose of the processing, the types of data, the categories of data subjects and the obligations and rights of the controller (Art. 28(1) and (3) GDPR). Changing provider does not continue the review chain but restarts it — classification under Art. 6(2) AI Act, DPIA assessment under Art. 35 GDPR, information under Art. 26(7) AI Act before putting into service.

Placing the 2025/26 BAG case law correctly

BAG 8 AZR 300/24 is an equal-pay judgment, not an algorithm judgment; transferring it to algorithmic selection is a trend, not a decided case. This distinction is essential for assessing risk.

By judgment of 23 October 2025 (8 AZR 300/24), the Federal Labour Court (BAG) held that a pairwise comparison with a single, highest-paid male colleague is already sufficient to trigger the presumption of discrimination under § 22 AGG and § 7 EntgTranspG respectively; a comparison with the median is not required. Previously, by judgment of 21 January 2021 (8 AZR 488/19), the BAG had applied the reversal of the burden of proof under § 22 AGG in the pay context.

Both decisions are equal-pay judgments, not algorithm judgments. The transfer is to be expected as a trend, not to be cited as one-to-one authority: if a single comparator case is already sufficient as an indication, there is much to suggest that a documented bias finding in a selection algorithm will all the more be treated as an indication within the meaning of § 22 AGG. This does not create a new catalogue of obligations, but a changed expectation as to how proof is furnished. For Austria this case law is not relevant: there, the labour and social court decides on the basis of the GlBG, with a different standard of proof.

What breaches cost — separated by area of law and country

The cost consequences run along three independent tracks: applicants' claims, fines under the AI Act and fines under the GDPR. They are not mutually exclusive.

Applicants' claims. In Germany, § 15(1) AGG gives compensation for the damage incurred, provided the employer is responsible for the breach of duty, and § 15(2) AGG gives appropriate monetary compensation for non-material damage; where a person is not hired, it may not exceed three monthly salaries if the person would not have been hired even in a selection free of disadvantage. As a practice value, compensation is often in the order of one to three gross monthly salaries per claimant (no statutory scale). Austria works with minimum amounts instead of a cap: under § 12 GlBG (Part I) and § 26 GlBG (Part II) respectively, compensation for pecuniary damage and compensation for the personal impairment are due — at least one month's pay where the position would have been obtained in a selection free of discrimination, and up to EUR 500 where the damage consists only in the application not being considered.

Fines. Art. 99(4) AI Act provides, among other things for breaches of the deployer obligations under Art. 26 and the transparency obligations under Art. 50, for fines of up to EUR 15,000,000 or up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher; breaches of the prohibited practices under Art. 5 are, under Art. 99(3), subject to up to EUR 35,000,000 or 7%. On the GDPR side, Art. 83(5) covers breaches of Art. 5, 6, 7, 9 and Art. 12 to 22 with up to EUR 20,000,000 or 4%, and Art. 83(4) covers, among other things, breaches of Art. 25 to 39 — including Art. 28 and Art. 35 — with up to EUR 10,000,000 or 2%. In Austria, § 62(1) DSG applies only subsidiarily: unless the act already constitutes an offence under Art. 83 GDPR or is subject to more severe penalties under other administrative penal provisions, it is to be penalised as an administrative offence with a fine of up to EUR 50,000 — covered there are, for example, intentional unlawful access to a data processing operation and the intentional transfer of data in breach of data secrecy (§ 6 DSG); the competent body is the Austrian Data Protection Authority (Datenschutzbehörde).