You can buy and run AI recruiting tools in the EU, but only if you can prove the controls behind them. A defensible purchase shows evidence for GDPR safeguards, EU AI Act high-risk duties, data-transfer mechanics, human review and worker-representative notice before the contract is signed, not after go-live. That evidence belongs in procurement, not in a demo.

The real problem in 2026 is the gap between what vendors say and what they can actually document. Almost every marketing page now mentions responsible AI, fairness and "GDPR compliant," yet the proof a buyer needs usually sits in a data room you have to ask for. Candidate sentiment adds pressure too: only a quarter of job applicants trust AI to assess them fairly. So unclear disclosure can cost you conversions long before a regulator ever shows up.

Which recruiting AI is high-risk?

Recruiting AI counts as high-risk the moment it materially shapes who gets seen, filtered, scored or evaluated. Under Annex III point 4(a), systems built for recruitment or selection land in the high-risk category when they place targeted job ads, analyse or filter applications, or evaluate candidates. The Commission's employment examples name concrete cases: job matching and ranking, sourcing across online platforms, scoring applicant answers and background-risk scoring all fall in scope.

What matters is the use and the influence, not the label on the box. A tool sold as a harmless "assistant" is high-risk if it ranks your shortlist, and any system that profiles candidates counts as high-risk no matter how narrow it looks. The deciding question is whether the system changes who enters or advances in the funnel.

The other side is just as important, so you don't treat every HR automation as a regulated system. Interview scheduling, narrow credential checks, employer-brand advertising and candidate-controlled CV help are named as outside scope or potentially exempt, because they don't evaluate or filter applicants. Sourcing and ranking, by contrast, never escape scrutiny when they decide who recruiters ever see. And that is exactly where well-meaning buyers tend to underclassify.

Which GDPR and DACH gates matter?

Before any AI screening runs, three gates have to be cleared: a lawful basis under GDPR, Article 22 protection against solely automated decisions, and a concrete processor contract with your vendor. These apply on top of the AI Act, not instead of it, and they are exactly the controls candidates and supervisory authorities can test directly.

GDPR duties before AI screening

GDPR Article 22 gives every applicant the right not to be subject to a decision based solely on automated processing with legal or similarly significant effects. The EDPB-endorsed guidance reads this as a general prohibition, not a later opt-out. A fully automated rejection is unlawful unless a narrow exception applies and you build in human intervention, the chance to express a view and a route to contest the outcome. Touch special-category data, and the bar climbs higher: now you need explicit consent or a substantial-public-interest basis.

Lawful basis and data minimisation sit underneath all of this, and the processor contract carries real weight. An Article 28 DPA has to be concrete, not a copy of the statute, because the EDPB controller-processor guidance is clear: a vendor's standard terms don't move responsibility off the controller. A data protection impact assessment becomes the practical home for documenting screening risk, and the AI Act expects you to feed the provider's instructions into that DPIA.

Note: EU hosting is a helpful signal, but never a complete transfer answer. After Schrems II invalidated Privacy Shield while leaving Standard Contractual Clauses conditionally valid, you still need to know which subprocessors touch candidate data, where support and admin access happens, and whether US access relies on the EU-US Data Privacy Framework, SCCs or another Chapter V route.

Betriebsrat, transfers and DACH specifics

In Germany, AI recruiting tools routinely trigger co-determination. Section 87(1) No. 6 BetrVG gives the works council a say over introducing and using technical devices meant to monitor employee behaviour or performance. §90 adds an information-and-consultation duty on work-process planning that explicitly names artificial intelligence. And when the works council has to assess an AI rollout, §80(3) supports bringing in an external expert.

The neighbouring markets work the same way with their own triggers. In Austria, ArbVG §96 requires works-council consent for control measures and technical systems that touch human dignity, and employers must let the council verify automated processing of staff data. Switzerland's revised FADP is technology-neutral and gives data subjects transparency plus a right to request human review of automated individual decisions.

For HR teams, the conclusion is simple: the question is no longer whether AI recruiting is allowed. It is whether your architecture produces the evidence, controls and audit trail to prove lawful use before candidates, regulators and the Betriebsrat ask. That shifts the buying decision from feature lists toward architecture-level proof, which is exactly where an EU-hosted, audit-ready recruiting platform earns its place.

When do AI Act hiring duties start?

The headline date for employment high-risk AI has moved: the original 2 August 2026 baseline gave way to a 7 May 2026 political agreement on the Digital Omnibus that shifts stand-alone high-risk areas, including employment, to 2 December 2027. Product-integrated high-risk systems move further out, to 2 August 2028. The Commission's announcement of the simplification package sets out the new sequencing. And because this still rests on a political agreement, it's worth confirming the formal adoption right before you act on any fixed date.

Much of the public content you'll read still quotes the older August 2026 deadline, so honestly, the timeline itself has become a place where outdated advice circulates. The Regulation entered into force on 1 August 2024 and remains a risk-based framework that lists CV-sorting software among high-risk recruitment uses. Penalties of up to €35 million or 7% of worldwide turnover sit behind the most serious infringements, with up to €15 million or 3% for other breaches. Still, the deadline matters more for your planning than the fine does.

The extra runway should change your sequencing, not your preparation. Inventory of current recruiting AI, use-case classification and early works-council involvement take months, not weeks, and none of it gets easier by waiting for the final text. Treat the later date as breathing room to build documentation properly, not as a reason to pause.

What evidence should AI vendors show?

Audit-ready vendors hand over a documented evidence pack, not a principles page. For high-risk recruiting AI, that pack maps directly onto the provider duties in the Regulation. A vendor who can't produce it is asking you to carry their compliance risk.

Data, bias and model provenance

The data story sits at the core. Article 10 data-governance requirements cover data origin, the original purpose of collection, how data was prepared and labelled, representativeness, bias examination, bias mitigation and known data gaps. Datasets have to be relevant, sufficiently representative and aware of the context they operate in. Where a vendor uses special-category data to detect or correct bias, Article 10(5) allows it only as an exception, and only with pseudonymisation, strict access control, no onward transfer, deletion rules and GDPR record-keeping.

Around the data, the technical file fills out the picture. Article 9 expects a continuous risk-management system, Article 11 a technical documentation set kept current, and Article 13 clear instructions for use that declare accuracy metrics, robustness, foreseeable risks and performance for specific groups. Article 15 requires declared accuracy alongside resilience against faults, feedback-loop bias and attacks such as data poisoning.

Oversight, logs and candidate rights

Human oversight is the line buyers most often wave through too quickly. Article 14 demands that overseers can understand the system's limits, spot anomalies, resist automation bias and actually override, reverse or stop an output. Oversight without authority, competence and an audit trail is decoration, not a control. Article 12 backs this with automatic event logging across the lifecycle, and as a deployer you keep those logs for at least six months.

Candidate-facing rights and contracts complete the pack. You want the Article 22 challenge and human-review workflow shown in practice, a concrete DPA with a full subprocessor map, and transfer documentation that names the mechanism behind any third-country access. None of this has to be public; the realistic standard is evidence in the procurement data room under NDA, and a vendor who refuses even that is telling you something.

What should HR ask AI vendors?

Take a fixed set of evidence-demanding questions into every vendor call, and treat a verbal "yes" as the start of the conversation, not the answer. The items below each ask for an artifact or a demo, drawn from the AI Act deployer and provider duties and the GDPR baseline.

Buyer-call items before signature

A good vendor call should end with evidence in the data room, not assurances in a slide deck. Teams that want AI sourcing speed without piling up compliance debt benchmark every provider against an EU-hosted, audit-ready recruiting architecture, then decide. That benchmark is exactly the role Atlas People-Search is built to play.

Which AI recruiting vendors are audit-ready?

Vendor transparency in 2026 is a test, not a leaderboard: some providers publish genuinely useful signals, but full Annex IV-grade evidence almost always lives in procurement materials, not on a public page. Greenhouse's public AI principles are a fair example of the better end, stating ISO 42001 AI-management certification, training of proprietary models only on anonymised, de-identified data such as location and time-to-hire, and a manual-review option for Talent Matching. Eightfold publishes a transparency checklist urging disclosure of data sources, training approaches and bias-test results, while Paradox commits to documentation principles at a higher level.

This is where Sprad's EU-hosted recruiting architecture serves as a reference point for buyers who want GDPR-oriented AI sourcing with documented data sources, Article 22 opt-out handling and Betriebsrat-readable audit support built in. We position Atlas People-Search as the architecture HR can benchmark against, and we expect to evidence these capabilities in procurement rather than ask you to take them on faith. Treat any capability claim, ours included, as something to verify against the document standard above before you sign.

A safer AI recruiting contract

HR genuinely needs AI recruiting capability. But the purchase that survives scrutiny is the one that produces evidence before regulators, candidates or the works council ever ask. High-risk classification, GDPR safeguards, the shifted timeline, the vendor evidence pack and the DACH audit trail all point to the same thing: buy the system you can document, and document it before you sign.

The next step is a short, ordered sequence. Inventory the recruiting AI you already run, classify each use case against Annex III, request the full evidence pack, pull Legal, IT and the Betriebsrat in early, and benchmark every shortlist vendor against an audit-ready EU architecture.

Frequently Asked Questions (FAQ)

Can an AI tool automatically reject candidates in the EU?

Generally no. GDPR Article 22 prohibits decisions based solely on automated processing that carry legal or similarly significant effects, and a rejection clearly qualifies. The narrow exceptions, such as explicit consent or a contractual or legal basis, still require safeguards: meaningful human intervention, the chance for the candidate to express a view, and a real route to contest the outcome.

Is AI active sourcing high-risk under the EU AI Act?

It depends on influence. Sourcing that materially shapes who gets seen or how candidates are ranked falls under Annex III point 4(a) as high-risk recruiting AI. Activity that is purely supportive or candidate-controlled, with no effect on employer selection, is generally treated as outside scope. The deciding factor is whether the system changes who enters or advances in your funnel.

Does EU hosting solve Schrems II for recruiting AI?

No. EU hosting helps but doesn't close the question. Schrems II left Standard Contractual Clauses conditionally valid while invalidating Privacy Shield, so you still need to confirm which subprocessors touch candidate data, where support and admin access occurs, which transfer mechanism applies to any third-country access, and whether supplementary measures and a transfer impact assessment exist.

What should the Betriebsrat be able to audit?

The works council should get an understandable audit trail, not raw model internals. That means the system's purpose, its monitoring functionality, access and log data, the human-review workflow, and clear worker-information materials. In Germany this flows from co-determination over technical monitoring devices under Section 87(1) No. 6 BetrVG, plus the consultation duties that explicitly cover AI in work-process planning.

Can vendors use sensitive data to test hiring bias?

Only exceptionally. AI Act Article 10(5) permits processing special-category data strictly for bias detection and correction, and only where no other data can achieve the goal. Buyers should request evidence of the legal basis, data minimisation, pseudonymisation, strict access controls and deletion, because under GDPR special-category processing otherwise needs explicit consent or a substantial-public-interest basis.