17 Nov 2026 EU AI Act compliance for HR software: what to test before December 2027
In short: EU AI Act compliance for HR software starts with one fact. AI that screens candidates or helps decide about workers is high-risk. The vendor must prove the system is accurate, fair, logged and under human control. The employer must use it as intended and keep watching it. Most of that proof comes from tests you repeat over time.
If your software ranks CVs, you’re building a high-risk AI system. The law names it in so many words. Most HR teams and HR-tech vendors know this by now. Fewer know what it means on a Tuesday morning: which tests to run, what to keep, and who in the chain has to do it. EU AI Act compliance for HR software comes down to evidence. Evidence takes months to build, so we’d start now rather than in 2027. What follows isn’t legal advice. Your legal team decides what applies. Your testers make the proof.
Is HR software high-risk under the AI Act?
Yes, when it uses AI for hiring or for decisions about people at work. The AI Act, the European law that sets rules for AI by risk, lists these uses in Annex III, point 4. The first group is recruitment and selection. Think of placing targeted job ads, filtering applications and scoring candidates. The second group is decisions about people already at work. That covers promotion, dismissal, handing out tasks based on someone’s behaviour or traits, and tracking how people perform. A tool that does one of these things with AI is a high-risk AI system, a system that’s allowed but has to meet strict rules. There’s a narrow way out. A system that only does a small task to prepare a decision, without real influence on the outcome, may not count. Don’t lean on that too fast: if it ranks or filters people, it shapes the outcome. One use is banned outright. Since February 2025, AI that reads emotions of workers is forbidden at work, except for medical or safety reasons. Our article on the EU AI Act for software teams covers the general rules. Here we stick to HR.
Who has to do what: the vendor or the employer?
The vendor carries most of the duties, but the employer can’t hide behind the vendor. The law calls the vendor the provider: the company that builds the system and sells it under its own name. The employer that uses it is the deployer. Both have their own list, and you need to know which side you’re on.
The vendor (provider)
The provider has to show that the system was built and tested to be accurate, fair and safe. That means a risk analysis, checks on the data, technical documentation, automatic logs and a way for people to oversee and overrule the system. After launch, the provider has to keep monitoring how the system performs, for its whole lifetime. That’s Article 72 of the AI Act.
The employer (deployer)
The deployer’s duties are in Article 26. Use the system as the vendor’s instructions say. Give human oversight to people who have the skills and the authority to step in. Make sure the data you feed it is relevant. Watch how it works, and report serious problems. Keep the logs for at least six months. And tell workers and their representatives before you use such a system at work. If you change the system a lot, or put your own name on it, you may become the provider yourself.
What should you test in AI for hiring and HR?
Test whether the system gets it right, for everyone, and whether a person can really step in. Software with AI inside doesn’t always give the same answer twice, so a single pass or fail proves little. You run the same checks many times and look at the pattern. These are the tests that turn the law into proof.
- Accuracy per group. Does the tool rank good candidates high, and is that true for women and men, older and younger people, people with a non-Dutch name? An average score can hide a group it fails.
- Bias. Swap one detail in an otherwise identical CV, like a name, an age or a gap for parental leave, and compare the outcome. Our article on AI bias testing explains how.
- Robustness. What happens with an odd CV layout, a scan, a CV in English, or a hidden line of text that tries to trick the AI?
- Human oversight. Can a recruiter see why a candidate scored low, and overrule it? Test that the button exists, that people use it, and that the override is saved.
- Logs. Check that every decision is logged with enough detail to explain it later.
Two tests sit on the employer’s side too. Does the tool still behave with your own vacancies and your own candidates? And do your recruiters treat the score as advice, or do they follow it blindly? You check that second one by talking to recruiters and looking at how often they disagree with the tool. AI can help a recruiter. The decision about a person stays with a person.
When do the rules apply, and why start now?
For hiring and HR, the high-risk rules apply from 2 December 2027. That date comes from the AI Omnibus, which took effect in July 2026 and moved the Annex III deadlines. Check the official text and the Commission’s pages before you plan, because dates have moved before. The ban on emotion recognition at work has applied since February 2025. And the GDPR, the EU privacy law, already gives candidates rights when a decision about them is fully automated.
December 2027 sounds far away. But the proof the law asks for is proof over time: test results from months of use, monitoring reports, logs. You can’t produce a year of results in the month before a deadline. Vendors also need time to fix what the tests find, and employers need time to switch tools if a vendor can’t deliver. A trade-off to name: if your tool only schedules interviews or answers practical questions, the high-risk rules may not apply at all. Check that first, before you build a test programme you don’t need.
What this means for your team
The first round of tests is the easy part. The hard part is the years after: the tool keeps changing, your vacancies change, and the law expects you to keep watching. Someone has to run the checks again after every update and follow up on what fails. They also have to report it so HR, legal and management can act. In most organisations that job doesn’t exist yet. We wrote about that ongoing loop in our article on AI monitoring after go-live.
For that ongoing work we built VeritasAI, our in-house tool. We use it to deliver AI testing as a service. VeritasAI keeps testing your AI application with real questions. An operational dashboard shows which questions fail and which law, regulation or framework each failure relates to. You see the prompt, the answer and a priority score. A management dashboard gives the overview a compliance officer asks for. Read more on our AI compliance testing page.
A first step for this week: ask your HR vendor, or your own product team, for the results of the last bias test. Ask when it ran and on which groups. If nobody can answer, you know where to start.
Build the proof the AI Act asks for while you build the product.
Frequently asked questions
Is a chatbot that answers candidates’ questions high-risk?
Usually not, as long as it only answers questions and doesn’t filter or score people. Candidates must be told they’re talking to an AI. Once it starts judging who goes through, the high-risk rules come in.
Does the AI Act ban AI in recruitment?
No. AI in recruitment is allowed, but as a high-risk use with strict rules. Only emotion recognition at work is banned, apart from medical or safety reasons.
Do we need to test if our HR vendor says the tool is compliant?
Yes, in part. The vendor proves the tool works as built. You’re responsible for how you use it: with your data, your vacancies and your people. Ask for the vendor’s test results and check the tool in your own setting.
What about the GDPR?
The GDPR applies next to the AI Act. Candidates have rights when a decision about them is made fully by a machine, including the right to have a person look at it. Involve your privacy officer early.