HR edition
Dutch decision-makers draw a line at dismissal and contracts, Workday must hand over more and more application data, and AI's top executives are themselves asking for the brakes to be applied.
Dutch organisations: dismissal and contracts stay human work
More than half of decision-makers think AI should not decide on its own about dismissal, contracts and pay. And in places where AI is allowed to, a human often only steps in once things have already gone wrong.
IT consultancy Dynamic People asked 317 IT and business decision-makers at medium and large Dutch organisations where they draw the line for AI that acts on its own. The result matters directly for HR. On dismissal, 51% say people must stay in control. For contracts with suppliers or clients, that's 53%. For pay and employment terms, 45%. People are strictest about money: 61% think AI should never decide on its own about financial actions above a certain amount. The most common limit is 10,000 euros (29%), followed by 50,000 euros (22%). 18% want no independent financial decisions by AI at all. The sharpest figure comes last: at organisations that do let AI act on its own, a human only gets involved in 79% of cases after damage has already been done. So that's already common practice at companies that have taken this step. For HR, this is a chance: write down now which decisions need a human signature, before a tool acts on its own. Not as a brake, but as an agreement you can point back to later. This week, write down which HR decisions will always stay human work at your organisation. AI that acts on its own works with permissions: you give a system access to, for example, a calendar, mailbox, personnel file or payment function, plus a goal. It then chooses the steps itself. Oversight can happen beforehand (approval at each step) or afterwards (checking what happened). Oversight beforehand takes time, so organisations often choose afterwards. That's exactly where the problem starts: a human only checks once the email has already been sent. What matters is permission management: who is allowed to act without approval? Setting everything to "human decides" sounds safe, but then you lose the time saved and oversight gets rushed. It's better to distinguish by consequence: something easy to undo (like scheduling an appointment) can be checked more lightly than something irreversible (like sending a rejection or changing an employment term). Keep in mind that the works council will ask: who decided this system was allowed to act on its own, and where is that decision recorded?
Background for subscribersThis edition shows that most organisations don't yet let AI decide on its own about dismissal, contracts or pay, but that oversight often only comes afterwards. The Workday lawsuit, the first AI Act inspections on CV screening, and Microsoft's requirements for Copilot users are all about the same thing: who checks an AI decision, and when. For you, this means you can write down this week exactly which HR decisions always need a human, and ask your suppliers whether they have a bias test and up-to-date documentation.
Workday case: 1.1 billion rejections on the table
The collective lawsuit against Workday over discrimination through AI screening is continuing. New court orders from late June and late July are forcing Workday to release more and more client data and bias-test data. Court documents show that since September 2020, 1.1 billion job applications have been rejected using its software. This involves alleged discrimination based on age, race and disability. Importantly: judges confirm that an employer and supplier can be jointly liable for a discriminatory outcome, even without intent. So don't blindly trust your supplier's assurances, check for yourself. This week, ask your ATS supplier in writing which parts use AI and whether a bias test has been done. Discrimination rarely comes from an explicit rule. More often, a model learns from past assumptions about who was "successful". If there was bias in that past, the model picks it up. It then finds features linked to that, such as a graduation year or place of residence. Age isn't a rule anywhere, but the effect is still there. That's why a bias test looks at outcomes, not code: you compare the chances of progressing to the next step for different groups of candidates. Checking this yourself requires data that HR often doesn't want to record, such as the age or background of rejected candidates. To see inequality, you have to measure it, and measuring touches on privacy. So measure at group level, not traceable to individuals, and discuss this beforehand with your privacy officer and works council.
Background for subscribersThis is a taster.
Subscribers read the whole HR edition: every story, each with its background in plain language.
Subscribe →Not ready for that? Read the free letter first →
AI and tech, made simple.
This is the HR edition. Subscribers get the background to every story, in plain language.
Subscribe
inklaretaal · Amsterdam · © 2026