AI in HR: tools that work, pitfalls to avoid

AI is changing HR from the ground up

One vacancy costs you six weeks of lead time, a few thousand euros in advertising and an unhealthy pile of emails asking "does Thursday at 14:30 work, or is Friday at 10:00 better?". In a small company with one HR person for 25 people, that hurts. Recruitment, onboarding, leave, reviews, contract changes, it stacks up, and the strategic work ends up on the "someday" list on Friday afternoon.

AI takes the sharp edges off, if you know where to apply it. Not by replacing the HR person (you cannot anyway, recruitment AI without human oversight is prohibited under the EU AI Act), but by removing the tedious steps in between. Below: where AI genuinely adds value for HR teams in 2025, where it goes wrong the moment you get lazy, and how to start without sending your works council into shock.

Where AI already makes a difference

Job ads that do not read like every other LinkedIn post

Nine out of ten LinkedIn vacancies open with "Are you the driven professional who...". That is not a vacancy, that is a template. AI solves this if you brief it properly: feed your brand voice, two customer stories and the real role profile into a Claude prompt and ask for three variants. Textio does this more systematically and flags male-coded language ("ambitious team of hunters"), but for a smaller company with four vacancies a year a good prompt library is enough. Time saved per vacancy: 60 to 80 percent, and response rates go up because your text finally stands out.

Screening and selection (the risky one)

This is where the bias bomb sits. HireVue analyses video interviews, Pymetrics uses cognitive games, Greenhouse and Workday have CV scoring built in, Recruitee and Homerun lag behind in the Netherlands. It works, provided you accept that the tool supports and does not decide. The recruiter stays responsible for rejections and has to be able to explain them. Anyone who gets lazy here and takes the shortlist at face value will sooner or later have a GDPR complaint on their desk.

Intake and scheduling, the real gold

This is where most time is saved and almost nobody talks about it. An intake bot that lets candidates ask questions about the role, salary range and location around the clock. An AI agent that schedules interviews on its own based on two manager calendars and the candidate's. A WhatsApp flow that reschedules cancellations. Not high risk, no works council drama, saves your HR colleague four hours a week. You build this with Cal.com plus a Claude agent, or a custom GPT trained on your own documentation. More applications on AI for HR.

Onboarding

AI chatbots answer a new employee's first 100 questions. From "where is the coffee machine" to "how do I request leave". Saves hours a week and gives newcomers an answer at the weekend too. Notion AI on top of a decent handbook works fine, as do the onboarding flows in BambooHR.

Predicting attrition

Predictive analytics indicate who is at risk of leaving. Not on gut feeling, but on patterns: fewer Slack messages, less response to feedback, more sick days, lower engagement in surveys. Workday Skills Cloud, Visier, Gallup Q12. For a smaller company a spreadsheet with rules is often enough, and that is fairer, because you still understand what is happening inside it.

Leave administration and routine communication

Leave requests over WhatsApp that land in the system automatically. Welcome emails, first-day information, mentor introductions. Periodic reporting to management on absence and contracts. The quiet time sinks where AI often has the biggest impact, and which are not "high risk" under the AI Act at all.

AI screens and sorts, people decide. Nobody wants to be rejected by an algorithm without an explanation.

The pitfalls

Bias in the data

If you train AI on historical recruitment data, you drag your own prejudices along with it. The well-known example: between 2014 and 2017 Amazon developed an internal AI recruiter and announced publicly in 2018 that the project had been shut down because the system systematically disadvantaged women, simply because the training data consisted mostly of files of male candidates hired in the past. The AI learned "successful candidate = man" and discriminated on that basis.

This is not a solved problem. Every recruitment AI system carries this risk. The solution is a combination of transparent criteria that you set in advance (rules you decide, not patterns the model invents), bias checks (monitoring dropout rates per group) and blind screening on name, photo and address in the first round.

Privacy and GDPR

Employee data is sensitive. Absence, reviews, salary, interview notes, those are GDPR categories you do not simply dump into a cloud AI. Make sure your tools are GDPR compliant, process data inside the EU only and come with standard data processing agreements. For the most sensitive parts (psychological profiles, interview data) on-premise or a private EU environment is worth serious consideration. Open to discussion per case.

High risk under the AI Act

Recruitment AI falls under the "high risk" category of the EU AI Act (Annex III). From 2 December 2027 that means mandatory transparency, human oversight, documentation, a fundamental rights impact assessment for public employers and the right for candidates to object. That date was postponed from 2 August 2026 in July 2026, the requirements themselves did not change. Not optional extras, legal requirements. Many existing AI recruitment tools do not meet them yet. Ask every supplier explicitly for documentation on AI Act compliance. If you get a vague email back, you know enough.

The human dimension

AI screens and sorts, people decide. Nobody wants to be hired by an algorithm, and nobody wants to be rejected by an algorithm without an explanation. Fixed rule: AI delivers the top 10 shortlist based on criteria set in advance, the recruiter decides which 5 get a call, the manager decides who gets hired. This is also the standard the AI Act prescribes.

Getting started with AI in HR: a practical plan

Month 1: diagnosis and goals Which HR processes cost you the most time right now? Job ads? CV screening? Leave administration? Which ones deliver variable quality (ask three new employees about their onboarding experience and you get three different stories)? Which are high risk under the AI Act and which are not? Set your goals: time per vacancy, time to hire, candidate satisfaction, diversity in the shortlist.

Month 2 to 3: pilot on a low-risk process Do not start with CV screening, that is high risk and calls for a works council process. Start with job ads, an onboarding bot or a scheduling agent. Tools: Claude or ChatGPT with well-considered prompts and a set of vacancy templates. Working method: AI produces a draft, HR reviews and adjusts. Measure progress weekly. If it leads nowhere after six weeks, stop and pick a different pilot.

Month 4 to 6: works council process for high-risk processes Only once the first pilots are running do you take on CV screening or attrition prediction. The works council has to be part of the decision here. Planning: briefing (1 session), demo (1 session), FAQ (1 round), formal request for consent (formal documentation). Allow plenty of time, a works council caught off guard says no.

Month 7 and beyond: scale up and integrate Extend successful pilots to other teams or vacancies. Update your standard HR processes, not as separate tools but as steps in workflows. Train new HR colleagues on the AI tools as part of their own onboarding.

How a project like this runs

I work in phases. First a short analysis of which HR processes deliver the most gain and which fall under a works council process. Then a targeted pilot on a low-risk process (vacancy content, onboarding checklist or scheduling bot). Only after that do I take on high-risk processes with a works council process. Between phases you evaluate whether to continue, so you do not get stuck in a six-month preparation track that nobody remembers the point of. I do a cost-benefit analysis up front.

The core of it

AI in HR works, provided you start with low-risk processes, take GDPR and the AI Act into account from day one, and communicate transparently with employees and candidates. Do not automate everything, discover step by step where AI adds value without the human dimension disappearing. The gain sits far less in clever CV screening than everyone thinks, and far more in job ads that do not sound like all the others, a scheduling bot that stops emailing about "Thursday or Friday", and an onboarding flow that works on a Sunday evening too.

Many companies start with job ads or an onboarding bot, and expand from there into leave administration and (after a works council process) CV support. No revolution, but more room for strategic HR work. That is where AI adds the most value in 2025.

Want to know which HR processes would deliver the most gain for your organisation? Take the free AI scan and get a personal analysis in your inbox within 1 minute.