AI in European recruiting after Annex III: the ATS conversation has changed

The Annex III line that runs straight through the ATS category

The EU AI Act's Annex III, the list of use cases the regulation classifies as high-risk, includes AI systems used for "recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates." The drafting is broad on purpose. An applicant tracking system that ranks candidates by algorithmic match to a job description falls inside the definition. A system that filters out applications below a confidence threshold falls inside it. A system that recommends which candidates a recruiter should contact first falls inside it.

What this means in practice is that European employers running these systems are deployers under Article 26 and providers of these systems are providers under Article 16. The obligations are substantial. Documentation, human oversight, accuracy and bias monitoring, transparency to candidates about the use of automated decision-making, and conformity assessment for the providers. The dates matter. The Annex III obligations were postponed by the Digital Omnibus political agreement of May 2026 to 2 December 2027, which gives the market sixteen additional months. The structural shape of the category has already shifted in advance.

What the European-built tools did

Teamtailor, the Swedish-headquartered ATS that has been one of the fastest-growing in European mid-market recruiting, made the most explicit structural response. The 2024 product release decoupled the candidate ranking and matching features from the core ATS so that an employer could disable the AI-driven ranking entirely without losing the underlying tracking and pipeline functionality. The implicit positioning was that the employer should make the deployment decision on the AI features deliberately rather than have them on by default. This is the kind of design choice that looks small in a product release note and is large in a procurement conversation with a DPO.

Recruitee, the Polish-and-Dutch ATS now part of the Tellent group, took a different path. The product retained its AI-driven candidate suggestions as a default but added detailed audit logs and explainability features that let an employer document the human oversight required under Article 26. The bet is that mid-market employers will not turn off the AI features because they are useful, but will need to defend the decision and the system's behaviour if a candidate complains or a regulator inquires.

SmartRecruiters, the American-headquartered enterprise ATS with significant European customer presence, sits in a different position. The product has the breadth and the integrations the European enterprise market expects, and the AI features are competitive. The structural question is the same one any American-headquartered AI provider faces under the EU AI Act: where is the conformity assessment, what is the human oversight architecture, and how does the system documentation map to the regulation's specific requirements. The vendors that have a clear answer to this have an advantage in the next eighteen months. The ones that do not have answers will lose deals to the ones who do.

Greenhouse, the American ATS that historically dominated American mid-market and enterprise recruiting and that has a meaningful European presence in tech companies, has moved more carefully on AI features than competitors. The conservative pace, often criticised in product reviews, looks more defensible against the AI Act timeline than the faster-shipping competitors that now have to retrofit their AI features for European compliance.

The works council factor

For German, French, and Dutch employers, the AI Act layers on top of the existing works council rights that already constrained AI deployment in HR. Section 87 of the German Works Constitution Act gives the works council co-determination rights on technical systems used to monitor employee behaviour, and ATS systems with AI ranking have triggered co-determination requirements in multiple cases since 2023. A 2024 ruling from the Federal Labour Court reinforced that the works council has a right to evaluate not just the ATS deployment but the model behaviour itself, which produced a wave of works council requests for documentation that vendors had not anticipated providing.

The practical effect is that a German mid-market employer considering an ATS in 2026 is making a procurement decision that has to clear two parallel review processes: the DPO's AI Act readiness review and the works council's co-determination consultation. The vendors that produce the documentation both processes need without a customer-specific commercial conversation are the ones who land deals faster. The vendors that treat documentation as a customer success conversation lose months on each enterprise procurement.

What the procurement question looks like now

The productive procurement question for a European HR team in 2026 is no longer "which ATS has the best AI features." It is "which ATS gives my organisation the cleanest path to Article 26 compliance, the cleanest works council consultation if applicable, and the cleanest defensibility to a candidate complaint about automated decision-making." The answer to that question is not obvious from the product demo. It requires reviewing the vendor's AI documentation, the model bias monitoring, the audit logs, and the conformity assessment.

A real case from the 2025 procurement cycle. A €400M German manufacturing company evaluated four ATS vendors over six months. The decision was made on documentation depth rather than feature parity. The chosen vendor, which was not the largest in the market, was the one whose AI documentation was specific enough that the works council could approve deployment in a single consultation rather than the three-meeting cycle the procurement team had budgeted for. The savings in implementation time exceeded the price differential between the vendor choices.

The category is sorting itself into vendors who treat the AI Act as a procurement asset and vendors who treat it as a compliance burden. The asset side is winning the deals the burden side is losing. The pattern is now visible enough that the next eighteen months will produce a clear before-and-after picture across the European mid-market.