The data is the problem
Every AI tool is only as good as the data that goes into it. That sounds like a no-brainer. But take a look at the average ATS database.
We connect with more than twenty different ATS systems. From Bullhorn to Carerix, from AFAS to Salesforce. What we encounter daily: job titles that are incorrect, skills that have never been filled in, outdated location data, and candidate profiles that haven't been updated in years.
That is the reality on which AI screening operates. Not the clean demo data you see at a trade show, but the messy reality of thousands of organizations using their ATS as a digital drawer where everything is thrown in.
So when you hear a supplier say "our AI matches candidates with 95% accuracy," my first question is always: compared to what? On their own cleaned dataset? Or on real production data where half of the fields are empty?
The arms race that no one can win
There is something funny going on when you take a step back.
Almost half of the applicants have ChatGPT polish their CV, rewrite their letter, and optimize their profile for the keywords that ATS systems look for. On the other hand, almost every employer uses automated systems to screen those same CVs.
Two AI systems fighting against each other. One optimizes to get through the filter, the other filters to pick the best. When you think about it, it's actually quite absurd. We have built a whole ecosystem where AI on one side makes texts look better than they are, and AI on the other side tries to see through that.
And then the other half. The candidates who just do it honestly. Who write their own letter, fill in their own CV. They are now being compared with polished AI versions and come out worse. Not because they are less good, but because their text looks less polished. The honest applicant is punished for not using AI.
That might be the most bizarre aspect of this whole development. We have created a system where you are almost forced to use AI, because otherwise you fall behind everyone who does. And meanwhile, no one trusts the outcome.
The fairness problem that no one expects
The promise of AI screening has always been objectivity. No human bias, purely based on qualifications.
But here it gets interesting. Researchers from VU Amsterdam and Stockholm University have followed an AI implementation at a large company for three years. What they discovered was something I didn't expect: the problem was not that the AI was unfair. The problem was that everyone had a different definition of "fair".
HR wanted consistency. The same rules for everyone, no exceptions. Managers wanted context. That one intern who doesn't match on paper, but you can just see is a star. The algorithm always chooses the HR definition, because consistency is easy to program. Context is not.
I think that is the most underestimated risk of AI in recruitment. Not that it makes mistakes. But that it enforces a very specific way of thinking and takes away your ability to use common sense.
And I haven't even mentioned the bias in the training data. Amazon had to scrap their AI recruitment tool years ago because the system had learned to score CVs with the word "women's" lower. This is often cited as an incident. It is not an incident. It is what happens when you train a system on how it has always been done.
Where it works
I don't want to give the impression that AI in recruitment is worthless. It is not.
Writing and optimizing job descriptions. Interview scheduling. Data cleaning. Sorting the first bulk on hard criteria. It really makes a difference.
The pattern is simple: AI is good at efficiency. Writing texts faster, scheduling faster, sorting faster. Where it doesn't work is at the selection decision itself. Harvard Business Review concluded earlier this year that there is simply no convincing evidence that AI selects better than existing methods. Not "AI doesn't work." But: "it doesn't work better than what we already had, and brings new risks with it."
That is quite an uncomfortable conclusion when you consider how much money is being pumped into AI recruitment.
The EU AI Act as a reality check
As of August this year, the high-risk requirements of the EU AI Act will be enforceable. Recruitment AI falls explicitly under that. What that means in practice: if you use AI tools for screening or ranking candidates, you must perform risk assessments, keep documentation, ensure human oversight, and register your system.
Many recruiters and HR managers I speak to are not yet dealing with this at all. They use AI tools from their ATS supplier or a standalone plugin, without knowing that they will soon be responsible for compliance themselves.
My advice: ask your supplier now how they are preparing. If they can't tell you exactly what they are doing about risk assessments and bias monitoring, then you have your answer.
Why we are building it anyway
We are not building Recruit Pivot because we think we have already solved the problem. We are building it because the current process is already broken without AI.
The average recruiter evaluates a CV in six seconds. There is nothing wrong with that, because after twenty years of experience, you have that feeling. But when you look at the rest of the process: the response time, the number of candidates who simply never hear anything, the quality of the data in the ATS. That is where the real gain lies.
But we are honest enough to say that we are still a long way from there. We know how difficult it is to keep bias out of your model. We know that data quality is the biggest obstacle, not the algorithm. And we know that a good hire depends on team dynamics, culture, growth potential, timing, and dozens of other factors that you cannot capture in a vector embedding.
What I want to leave you with
If you are considering AI in your recruitment process today:
Be critical of what suppliers promise. Ask for independent evidence, not their own case studies. If they claim that their AI matches "95% accurately," ask precisely compared to what.
Use AI where it has proven to work. Efficiency, not selection. Let it write texts, schedule interviews, clean data. But let a human make the final decision.
And prepare for the EU AI Act. August 2026 is in four months.
Distrust anyone who claims they have already solved the problem. Including us. We are on the way, but honest about how far we still have to go.