AI Inventory in HR: 6 Typical Tools and Risk Classification

Main article: AI Inventory in 8 Steps — how to record and classify AI systems and assign the provider or deployer role.

The six profiles each classify one typical HR AI tool: the risk class under the AI Act, the resulting obligations with their legal reference, and the distinction from the neighbouring case. What matters is the function of the system (Art. 6(2) and (3) AI Act), not the product name — the same product can be high risk or limited risk depending on its configuration. The product examples given reflect the position as at 2 May 2026. The general date of application of the AI Act under Art. 113 (2 August 2026) has been reached.

1. AI recruiting (Talent Scout, HireVue)

AI systems that assess, rank or pre-select applications are high-risk AI: Annex III no. 4 of the AI Act covers the area of employment, workers management and access to self-employment. Risk: HIGH (Annex III, 4). Obligations: deployer obligations under Art. 26 AI Act, for Annex III systems from 2 December 2027 (Regulation (EU) 2026/1744) — use in accordance with the instructions for use (paragraph 1), human oversight by competent persons (paragraph 2), retention of logs for at least 6 months (paragraph 6), information of workers' representatives and affected employees before putting into service (paragraph 7), information of affected persons, e.g. applicants (paragraph 11) — plus, as a practical recommendation, bias testing and algorithm audit documentation; the seven components we recommend for a bias assessment are set out in the FAQ "What must an HR AI bias assessment contain as a minimum?". A fundamental rights impact assessment (FRIA, Art. 27) is not required here: Art. 27(1) only obliges bodies governed by public law, private entities providing public services, and deployers under Annex III no. 5(b) and (c) (for whom Art. 27, as part of Chapter III Section 3, applies to Annex III systems from 2 December 2027).

What to do: Record for each recruiting system which output feeds into the selection decision — score, ranking, pre-selection or text summary — and check the relevance of the input data for the intended purpose (Art. 26(4) AI Act). Practical recommendation: include the information for applicants in the job advertisement and the privacy notice, not only in the rejection letter.

Where the exception ends: Under Art. 6(3) AI Act, a system listed in Annex III is exceptionally not considered high risk where it does not pose a significant risk to health, safety or fundamental rights — but this carve-out does not apply where the system performs profiling of natural persons. The exception requires that the system performs only a narrowly limited preparatory task, a limited task within the use case, or a purely supporting activity. Where a system produces recommendations or assessments capable of materially influencing human decisions, it remains high-risk AI. Practical recommendation: in case of doubt, classify as high risk and record the reasoning in the inventory.

2. ATS with ML score (Personio, Workday)

An applicant tracking system becomes high-risk AI the moment its score assesses, ranks or filters out applications; pure data extraction without assessment logic is not covered. Risk: HIGH where there is automatic filtering out. Obligations: where classified as high risk, the full deployer obligations under Art. 26 AI Act (listed in section 1), plus a human final decision and a privacy notice for applicants.

The dividing line in detail: As soon as CV content is evaluated (match score, recommendation ranking), Annex III no. 4 applies. Mere pre-sorting of the inbox without assessment logic is unlikely to be high risk. Further borderline cases that are not high risk: see the FAQ on the exception under Art. 6(3).

What to do: Compare the provider's description of the score with actual use — what matters is what your HR department does with the output. Practical recommendation: document for each selection round who made the final decision and whether the score was departed from. Where the ATS provider processes applicant data on your behalf, a data processing agreement is also required (section "Data protection obligations alongside the AI Act").

3. Video interview AI (with emotion detection)

Video interview systems with emotion recognition are in principle not usable in the employment context. Risk: partly prohibited (Art. 5). AI systems to infer emotions in the workplace are prohibited under Art. 5(1)(f) — irrespective of the intended purpose; the only exception is for medical or safety reasons.

The prohibition has applied since 2 February 2025 (Art. 113 AI Act, Chapters I and II). Three further prohibited practices may be engaged by sourcing and video tools: social scoring by employers (Art. 5(1)(c)), the untargeted scraping of facial images from the internet or from CCTV footage (point (e)) and biometric categorisation to infer sensitive attributes (point (g)). For fields of use outside the prohibition, Art. 50(3) AI Act additionally requires deployers of emotion recognition and biometric categorisation systems to inform the persons concerned.

What to do: Obtain written confirmation from the provider as to which analysis functions the system contains, and document in the inventory that the emotion and biometric analysis functions have been switched off or contractually excluded. Where biometric data are processed for the purpose of unique identification, the processing prohibition of Art. 9(1) GDPR also applies, with the narrow exceptions of paragraph 2. Practical recommendation: repeat this check before every contract renewal.

4. Employee performance tools

Performance tools are high-risk AI as soon as they assess employees in an automated way or generate promotion or dismissal recommendations. Risk: HIGH where there is automated assessment. Obligations: deployer obligations under Art. 26 AI Act as in section 1, transparency, right to object, and where applicable a DPIA under Art. 35 GDPR — a FRIA under Art. 27 does not apply to private employers that do not provide public services (section 1).

Employee monitoring also falls under Annex III no. 4. A pure analytics dashboard without assessment of individual employees must be distinguished from this — the distinction belongs in the inventory together with its reasoning.

What to do: Information of workers' representatives and of the affected employees before putting into service (Art. 26(7) AI Act); in Germany, co-determination of the works council under § 87 BetrVG applies in addition, in Austria co-determination under the Labour Constitution Act (ArbVG §§ 96/97) — details in the section "DE/AT". Practical recommendation: name an escalation route, with name and function, through which employees can object to an automatically generated assessment.

5. Pay analysis tools (pay-gap detection)

Pay analysis tools that evaluate pay gaps at group level without assessing individual employees are not a high-risk case — but they belong in the AI inventory in full. Risk: minimal to limited, as long as no assessment of individual employees takes place. Obligations: no high-risk obligations under Art. 26 AI Act; as a practical recommendation, an entry in the AI inventory with reasoning for the classification, plus an internal acceptable use policy setting out the permitted levels of analysis.

Where the classification tips: Where the tool generates assessments or recommendations about individual employees — for example on pay adjustment or promotion — that are capable of materially influencing human decisions, the classification under Annex III no. 4 must be reassessed.

The substantive law behind it (Germany): § 7 of the German Pay Transparency Act (EntgTranspG) prohibits agreeing or paying lower pay on grounds of sex for equal or equivalent work; § 3 EntgTranspG prohibits direct and indirect pay discrimination in respect of all pay components. The individual right to information under §§ 10 to 12 EntgTranspG exists in establishments with, as a rule, more than 200 employees; the comparative pay (statistical median) need not be stated where the comparable activity is performed by fewer than six employees of the other sex.

Austria: § 11 GlBG requires company classification rules and norms of collective legal regulation to observe the principle of equal pay for equal or equivalent work and prohibits criteria leading to discrimination; § 3 no. 2 GlBG prohibits discrimination in the determination of pay.

EU level (directive addressed to the Member States): Art. 7 of the Pay Transparency Directive (Directive (EU) 2023/970) provides that employees may request information about their individual pay level and about the average pay levels, broken down by sex, of the categories of employees performing equal or equivalent work, and that the employer replies in writing within two months.

6. Chatbot for employee questions (onboarding)

An HR chatbot that only provides information is not high-risk AI — but it is subject to the transparency obligation. Risk: limited. Obligations: bot notice under Art. 50.

Art. 50(1) AI Act requires AI systems intended to interact directly with natural persons to inform those persons that they are interacting with an AI system; the addressees there are the providers. Deployer obligations are set out in Art. 50(3) (emotion recognition, biometric categorisation) and paragraph 4 (deepfakes); where the system generates synthetic audio, image, video or text content, the marking obligation of Art. 50(2) applies in addition, which likewise falls on the provider. Practical recommendation: check the bot notice yourself before rollout, even though the obligation falls on the provider.

Where the role changes: Anyone who substantially modifies the underlying model or places the system on the market under their own name becomes a provider under Art. 25 AI Act. For everyone working with the bot, the AI literacy obligation under Art. 4 AI Act applies. As soon as the bot assesses employees or applicants instead of merely providing information, the classification under Annex III no. 4 must be reassessed.

Core fields: how to document an HR AI tool in the inventory

As a practical recommendation — there is no statutory list of fields — an inventory entry for each HR AI system records six fields: name and provider, use case, role, risk class, data types and oversight. Only these six fields make the set of obligations per system decidable; a seventh field records the training status for the AI literacy obligation.

• Name and provider: product, manufacturer and contracting party — the basis for clarifying the roles of provider (Art. 3 no. 3 AI Act) and deployer (Art. 3 no. 4 AI Act).
• Use case: what the system does, which data go in and which outputs come out — the point of reference for Annex III (Art. 6(2) AI Act).
• Role: provider, deployer or both; change of role in the event of a substantial modification (Art. 25 AI Act).
• Risk class: prohibited (Art. 5 AI Act), high risk (Annex III), GPAI or minimal — and, if you rely on Art. 6(3) AI Act, with written reasoning in the inventory (practical recommendation).
• Data types: which personal data are processed and whether special categories are affected (Art. 9(1) GDPR).
• Oversight: who monitors the system and how escalation works (Art. 26(2) AI Act).
• Training status: who operates the system, when which training was completed and where the evidence is held — the basis for the AI literacy obligation (Art. 4 AI Act, applicable since 2 February 2025) (practical recommendation; Art. 4 AI Act requires AI literacy itself, not this field).

For high-risk systems, the evidence from section 1 is added: retention of logs, relevance check of the input data, and information of workers' representatives and of the affected employees.

Provider or deployer: who bears which obligation?

In HR, companies are as a rule deployers, not providers — the role is determined by Art. 3 no. 3 and no. 4 AI Act and can switch under Art. 25. Providers develop or distribute the tool and bear the product obligations; deployers use it under their own responsibility and bear the obligations of use. The following allocation applies to high-risk AI systems under Annex III; where the risk is minimal or limited, these obligations do not apply — with the exception of provider registration under Art. 49(2) AI Act in the case of a self-assessment under Art. 6(3).

On the EU database in detail: the provider obligation under Art. 49(1) AI Act applies to all high-risk AI systems under Annex III with the exception of the systems under no. 2 (critical infrastructure) — it is limited neither to public bodies nor to Annex III no. 1. Under Art. 49(2), providers that conclude under Art. 6(3) that their system is not high risk must also register.

Data protection obligations alongside the AI Act

The AI Act does not replace the GDPR: every HR AI that processes personal data additionally needs a legal basis under Art. 6(1) GDPR — processing is lawful only if at least one of the six grounds is met.

• Special categories: health data, information on ethnic origin, on religious or philosophical beliefs and on trade union membership, as well as biometric data for the purpose of unique identification, are subject to the processing prohibition of Art. 9(1) GDPR, with the narrow exceptions of paragraph 2.
• Data protection impact assessment: Art. 35(1) GDPR requires one in advance where a high risk is likely, in particular in the case of new technologies; paragraph 3 names as a standard case the systematic and extensive evaluation of personal aspects (profiling) on which decisions producing legal effects are based.
• Processing on behalf of a controller: under Art. 28(1) GDPR the controller may only work with processors providing sufficient guarantees of appropriate technical and organisational measures; paragraph 3 sets out the minimum content of the contract.
• Requests from applicants and employees: requests under Art. 15 to 22 GDPR must be responded to without undue delay and at the latest within one month; where the extension is necessary because of the complexity and number of the requests, the period may be extended by two months, provided the person is informed within one month, with reasons (Art. 12(3) GDPR).
• Data protection officer Germany: from, as a rule, at least 20 persons constantly engaged in automated processing — and irrespective of the number of persons where processing operations are subject to a data protection impact assessment under Art. 35 GDPR (§ 38(1) BDSG).
• Data protection officer Austria: no numerical threshold going beyond Art. 37 GDPR; the obligation to designate follows directly from Art. 37 GDPR, while § 5 DSG additionally governs confidentiality and freedom from instructions in the public sector.

Supervision: in Germany, the authorities competent under Land law supervise non-public bodies (§ 40(1) BDSG); in Austria, the Austrian Data Protection Authority (Datenschutzbehörde) is the national supervisory authority (§ 18(1) DSG).

Deadlines: what applies on 16 August 2026

The dates under the AI Act relevant to HR AI have largely been reached. All dates in the table are dates of application of the AI Act under Art. 113; the status column refers to 16 August 2026.

The Digital Omnibus Regulation (EU) 2026/1744 has been in force since 27 July 2026 and has postponed the Annex III deadline: the high-risk obligations of Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027 and to Annex I systems from 2 August 2028; details in the FAQ "What does the Digital Omnibus Regulation change for HR AI?".

DE/AT: the national framework alongside the AI Act

Alongside the AI Act, Germany and Austria differ in the accompanying national framework for protection against discrimination and workplace co-determination — for HR AI the decisive points are above all the evidentiary rule in a discrimination case, the legal consequence and the co-determination provision; the table adds the time limit, job advertisement rules and contact point. The AI Act itself (Art. 5, Art. 10, Art. 26, Art. 27, Annex III no. 4) applies identically in both states (EU Regulation 2024/1689).

The protected characteristics largely coincide — § 1 AGG names race and ethnic origin, sex, religion or belief, disability, age and sexual identity; in Austria, § 3 GlBG prohibits discrimination on grounds of sex, and § 17 GlBG discrimination on grounds of ethnic affiliation, religion or belief, age or sexual orientation. What differs are the evidentiary rule, legal consequence, time limit, job advertisement rules, contact point and co-determination:

On prevention, German law imposes an organisational duty of its own on the employer: under § 12(1) to (3) AGG, the necessary measures to protect against discrimination, including preventive measures, must be taken; suitable training of employees counts as fulfilment of this duty. At EU level, Art. 5(3) of Directive (EU) 2023/970 requires job advertisements and job titles to be gender-neutral and recruitment procedures to be conducted in a non-discriminatory manner.

Fines and sanctions for HR AI

The framework depends on the type of infringement, not on the tool: prohibited practices are at the top, deployer obligations one level below, and data protection infringements follow their own catalogue.

Two clarifications: for SMEs and start-ups, the lower of the two values applies under Art. 99(6) AI Act; for small mid-caps, the corresponding cap in Art. 99(6a) AI Act has applied since Regulation (EU) 2026/1744 (in force since 27 July 2026). And the AI literacy obligation of Art. 4 AI Act does not fall under Art. 99(4) — there is no separate EU fine range for infringements of the AI literacy obligation; sanctions follow national law within the framework of Art. 99(1), in Germany since 29 July 2026 under the German AI Market Surveillance and Innovation Promotion Act (KI-MIG). Its own fine provisions in § 15(1) to (3) do not cover Art. 4 AI Act. Alongside the fines there is the civil law side: in Germany, damages and compensation under § 15 AGG, in Austria under § 12 and § 26 GlBG.