Not every AI recruiting tool needs works council approval in Germany. Approval is likely where the system monitors employees or establishes binding, recurring rules for hiring decisions. In other cases, the employer may primarily owe timely information and consultation – but that is still a real obligation, not a courtesy meeting.

A German works council, or Betriebsrat, is an elected employee-representation body governed by the Works Constitution Act. This article is practical orientation, not legal advice. The exact result depends on the tool’s purpose, settings, data flows and the structure of the German establishment, so employers should obtain employment-law and data-protection advice for their implementation.

Start with the legal question, not the label “AI”

The most useful test is functional. Does the system turn candidate data into a score, ranking or exclusion? Does it make the work, behaviour or output of recruiters measurable? Does the organisation intend to reuse those rules whenever it hires? Those three questions reveal more than whether a vendor calls its product a copilot, an assistant or an automated workflow.

The German statute sets out two central routes. Under section 87(1)(6) of the Works Constitution Act, the works council co-determines the introduction and use of technical facilities intended to monitor employees’ behaviour or performance. Under section 95, personnel-selection guidelines for hiring and other personnel measures require its consent; section 95(2a) expressly says that this remains true when artificial intelligence is used to create those guidelines.

Employee monitoring: section 87(1)(6)

Candidate analysis alone does not automatically make an AI tool an employee-monitoring system. Applicants are not yet the employer’s employees for the purpose of this provision. The issue changes when a recruiting system records, compares or exposes individual recruiter activity: interview duration, response times, approval behaviour, output targets or similar traces that can be used to evaluate a member of staff.

The key is the system design. A dashboard designed only for aggregate service management is not the same as a dashboard that lets managers drill into named recruiters and measure them over time. Employers should show the works council precisely which logs exist, who can access them, whether identities can be masked and whether any data can be exported for performance management. Those details are usually more important than a generic assurance that no one intends to monitor anyone.

Selection guidelines: section 95

A selection guideline is, in practical terms, a repeatable rule that guides personnel selection. An AI configuration can become one when it defines mandatory criteria, fixed weightings, threshold scores or automatic rejection logic for applicants. A rule such as “candidates below this score do not proceed” is very different from a structured summary that a trained recruiter assesses independently against documented, open criteria.

Calling an output a recommendation is not conclusive if hiring teams follow it as a rule in practice. Conversely, meaningful human review can reduce the automation risk, but it needs to be real: the reviewer needs access to relevant context, the authority to override the output and a record of why a recommendation was accepted or rejected. This distinction should be made before the procurement decision, not after the system has been configured.

When is information and consultation required instead of consent?

Even where sections 87 and 95 do not create a consent requirement, section 90 of the Works Constitution Act requires the employer to inform the works council in good time, with the necessary documents, about planned work processes and workflows including the use of artificial intelligence. The employer must discuss the planned measures and their effects on employees early enough for council proposals and concerns to be considered.

This matters for a candidate-facing tool because it can still change the tasks, skills, workload and accountability of recruiters. The works council may also bring in an expert where it needs to assess the introduction or use of AI; the Act treats that assistance as necessary in that context. Do not confuse this information right with a blanket product veto, but do not treat it as optional either.

The EU AI Act adds another information obligation for high-risk workplace systems. Its Annex III lists AI intended for recruitment or selection, including filtering applications and evaluating candidates, among the high-risk areas. Under Article 26(7), employers deploying a high-risk AI system at work must inform workers’ representatives and affected workers before use, where applicable under the relevant national procedures.

A workable AI works agreement is specific enough to audit

A works agreement is a written agreement between the employer and works council that governs a workplace matter. For recruiting AI, its purpose clause should say what the tool is permitted to do – for example structure interview information or support a human screening step – and what it must not do. “Use AI lawfully” is too vague to allow either side to test compliance.

The agreement should then connect the permitted purpose to the data. It should identify the candidate fields the tool may process, the employee-use data that may arise, the systems and service providers involved, and data that is off limits. It should set a documented retention approach: which records are needed for which purpose, who owns deletion and what happens to logs, exports and backups when the purpose ends. A broad promise to retain data for future model improvement is not a defined retention purpose.

It should also contain a clear ban on secondary performance analysis. Recruiting-tool data must not become a hidden scorecard for individual recruiters or be reused in unrelated personnel decisions. Human review, role-based access, audit logs, change management and supplier documentation belong in the same operating model. So does an objection and escalation route: it should say who can flag an incorrect or impermissible use, who can pause the workflow, how an override is recorded and how recurring issues are reviewed. That is more useful than an abstract promise of “human oversight”.

Plan the rollout as a governance workstream

The law does not prescribe a fixed negotiation period for an AI works agreement. For planning purposes, employers should budget in weeks rather than days, and should expect months where the system uses binding scores, extensive integrations, voice or video analysis, or raises disputed monitoring questions. This is a planning rule rather than a statutory benchmark: a launch date should follow the assessment of the actual scope, not force it.

A robust sequence begins with a process map. The employer identifies every input, output, decision point, user role, integration and log. The next step is a short, intelligible evidence pack for the works council and privacy team: a live demonstration, data-flow diagram, sample outputs, retention design, access model, supplier instructions and an explanation of known limitations. That joint review should lead into the agreement, a limited test under the agreed controls, and a scheduled review before wider deployment.

Preparation prevents most avoidable friction. Employers should be able to show where human decision-makers intervene, how they can override the system, how candidates receive transparent information, and how the organisation will stop or roll back a workflow. It also helps to distinguish the workflow under discussion: AI interview and voice recruiting raises different operational questions from document-led screening, while both need a clear account of data and decision impact.

GDPR and the EU AI Act are separate layers

Works council involvement does not replace data protection. The GDPR requires lawful, transparent and purpose-limited processing. Article 22 gives people protection against decisions based solely on automated processing when those decisions have legal or similarly significant effects. A reviewer who only rubber-stamps a model output may not provide meaningful human involvement. Where the planned processing is likely to create a high risk to people’s rights and freedoms, the organisation must assess whether a data protection impact assessment is required under Article 35 GDPR.

As of 20 August 2026, the official text of the EU AI Act applies in general from 2 August 2026. Recruitment and selection systems are listed in Annex III as high-risk AI; pre-existing high-risk systems have a specific transition rule where they undergo significant design changes after that date. Employers should therefore obtain and assess the provider’s instructions for use, intended purpose, human-oversight measures and relevant technical documentation instead of relying on a simple “AI Act ready” statement.

For organisations operating across Europe and the United States, the works-council question is specifically German: there is no equivalent universal works council approval rule in US employment law. Yet an employer with a German establishment still needs to assess German participation rights for that establishment, while GDPR and AI Act obligations may apply to the relevant EU processing and use. One global procurement decision should not flatten those distinct legal layers.

Tool choice cannot solve the governance question by itself

The product category can help a team make its intended use more concrete. A voice interview workflow, for example, calls for a clear account of prompts, recordings, scoring, human review and candidate communication. A CV-screening workflow instead puts particular focus on the data fields, filtering criteria and treatment of rankings. Neither label decides the legal outcome; the configured process does.

The same applies when evaluating providers. Compare the controls that matter to the deployment: whether rules can be changed, what logs are available, whether rankings can be switched off, how deletion works and which outputs reach hiring managers. The overview of AI interview and voice tools is a useful starting point for product research, but it cannot replace the organisation’s own works-council, privacy and AI-governance assessment.

An important limit to state openly

A works agreement can govern the employer’s and works council’s operating rules. It does not automatically create a legal basis for every type of processing, remove discrimination risk, validate a solely automated decision or certify a system’s AI Act compliance. That limit is precisely why the works council, data-protection team and recruiting owners should work from the same process map.

FAQ: works council approval for AI recruiting

Does every AI applicant-screening tool need works council consent?

No. Consent depends on the function, especially whether the tool creates selection guidelines under section 95 or is a technical facility intended to monitor employees under section 87(1)(6). Early information and consultation under section 90 may still be required.

Can a human click make an automated recruiting decision compliant?

Not by itself. The person must be able and authorised to understand, challenge and override the output using relevant context. A nominal approval step that always follows a score is not a reliable control.

Can a works agreement replace GDPR compliance?

No. A works agreement can define operational safeguards, but lawful basis, transparency, data minimisation, retention and data-subject rights remain separate GDPR questions. A high-risk processing operation may also need a data protection impact assessment.

Is AI recruiting prohibited by the EU AI Act?

No. Recruitment and selection AI is identified as a high-risk area, not universally prohibited. That classification brings obligations that depend on the role of the provider and the employer deploying the system.