AI in hiring, promotion and assessment: where high-risk duties apply from 2 Dec 2027.

AI in HR is moving from writing job ads to judging people's work. In Germany the share of companies using AI to assess employee workload more than doubled in two years, from 6% in 2024 to 13% in 2026, and 12% already use it to assess performance, according to Bitkom's August 2026 survey of 602 firms. In the United States, 39% of organisations have implemented AI in HR, with recruiting the leading use at 27% (SHRM, April 2026).

Each step brings AI closer to decisions for which the EU sets high-risk duties. Recruitment, selection, promotion, task allocation and performance evaluation are listed in Annex III of the AI Act. Under Regulation (EU) 2026/1744, the duties for those uses apply from 2 Dec 2027. That gives HR leaders 13 months from today to know which tools they run, what they decide and who answers for them.

A market built around hiring

The HR & Workforce brief places 28 AI offerings and two provider records (Workday, Rippling) in HR & Recruiting; 15 of the 28 offerings serve recruiting and hiring. They range from job-ad writers and sourcing agents to interview note-takers and autonomous screening agents: Ashby AI, Greenhouse AI, HireVue, iCIMS Copilot, Metaview, Paradox, Phenom, Retorio, SeekOut, softgarden, Teamtailor Co-pilot, Textio, Winston from SmartRecruiters and Workable's recruiting agents. The other half are HR suites with built-in AI, such as Workday, SAP SuccessFactors, Oracle HCM, Personio, Rippling and Cegid, plus talent intelligence, learning, performance and an employee service agent.

The concentration in recruiting makes sense. Application volumes are high, the work is repetitive, and the business case is easy to state. It is also where the evidence is strongest in both directions.

+18% job starts when AI led structured interviews, in a field experiment with about 70,000 applicants (Erasmus University, Oct 2025).

What the evidence shows

On the upside, a field experiment with about 70,000 applicants found that AI-led job interviews raised job starts by 18%, with higher offer rates and better 30-day retention. On the downside, peer-reviewed research has shown name-based bias in model-based CV ranking, and enforcement of disclosure rules has been weak. New York's Local Law 144 requires bias audits of automated hiring tools, yet the State Comptroller's audit found two complaints in two years and far more potential non-compliance than the city's own review.

Both findings point the same way. AI can improve hiring when the process is structured, validated and audited before and after go-live. It cannot be bought as a black box and trusted because the vendor says it is fair.

2 Dec 2027 is the date from which EU high-risk duties apply to AI used in recruitment, selection, promotion and performance evaluation.

What providers publish about your people's data

ETHORITY's Good AI Prism research read the pricing, security and legal pages of all 30 entries in October 2026. HR is better documented than most areas on hosting: 19 of 30 document an EU hosting option. Personio states that assistant data is stored in Frankfurt and never leaves the EU. Teamtailor lets customers choose between Ireland and Oregon. Ashby states that data is stored in the United States, with EU transfers under the Data Privacy Framework or standard clauses.

Training terms matter more here than anywhere, because the data describes candidates and employees. Eighteen of 30 document that customer data is not used for training, or only with opt-in. Two state otherwise. Workday says it automatically enables data contributions for AI features in purchased products, with an opt-out. SmartRecruiters reserves the right to use anonymised or pseudonymised data to develop and train AI features for Winston. Neither is unlawful in itself. Both are decisions an employer should take consciously, with the works council where one exists.

Prices are rare

Prices are rare. Six of 30 publish one, mostly the newer recruiting tools: Job Studio charges per active job, Ashby and SeekOut publish monthly plans, 360Learning lists a per-user price. The large suites quote.

The governance gap inside HR

SHRM's survey adds an uncomfortable detail: 52% of organisations do not involve HR in their AI strategy, and only 16% of HR functions use ROI as a success measure. HR is asked to deploy tools that touch every employee, while the strategy is set elsewhere and the value is rarely measured.

In Germany the gap is legal as well as organisational. Under the Works Constitution Act, technical systems that can monitor performance or behaviour require co-determination, and employers must inform the works council about planned AI use. A tool that scores workload is exactly such a system. The practical answer is to bring the works council in at the shortlist, not after the contract.

Use case. An interview note-taker decides nothing on its own, but its summaries feed a hiring decision. Treat it as part of that process: record which interviews it covers, who reads the summaries and who decides, and show the works council before go-live.

The quieter half: service, learning and literacy

Not every HR use carries high-risk duties. The most common uses in the Bitkom survey are supportive: individual training at 16% of firms, up from 12% in 2024, work references and onboarding at 14% each, and answers to internal HR questions at 13%. These uses carry lighter duties and often faster returns, because they take routine work off HR teams without deciding anything about a person.

They still carry one duty that already applies. Since 2 Feb 2025, Article 4 of the AI Act requires employers that deploy AI systems to ensure a sufficient level of AI literacy among the staff who use them. HR usually owns that programme for the whole organisation. The practical form is simple: a record of who uses which system, what training they received, and when. It also tends to surface the tools nobody had listed.

What high-risk duties will ask of you

For high-risk uses, deployers must use systems according to the provider's instructions, assign human oversight to competent people, monitor operation, keep logs, and inform workers and their representatives before use. Providers must supply documentation, accuracy information and the basis for conformity. None of this is new to a well-run HR function. What is new is that it must be shown, per system, with evidence.

That makes the inventory the first deliverable. Many HR teams cannot list every AI feature switched on in their suite, let alone the tools recruiters bought on a card. The Bitkom data on rising assessment use suggests the list is longer than most leaders think.

What to do Monday

1. List every AI feature in HR. Include suite features switched on by default and tools bought by individual recruiters.
2. Classify each by intended use. Recruiting, promotion, allocation and evaluation are Annex III uses; mark them for 2 Dec 2027.
3. Check training defaults. Ask each provider in writing whether candidate or employee data trains its models, and how to switch it off.
4. Bring in the works council now. Share the inventory and agree how assessment tools will be introduced and reviewed.
5. Set one measured outcome. Time to hire, quality of hire or retention at 90 days, with a baseline before go-live.

The Legal & Compliance topic area covers the wider AI Act calendar, and the HR & Recruiting area map holds all 30 profiles.