If Nobody Can Say Who Rejected a Candidate, TA Owns That Risk

Picture the scenario Aye Kalenok, founder and CEO of Kala Talent, lays out in a recent op-ed for Mexico Business News: a screening system scores one candidate 82 and another 61. The recruiter interviews the first. The second application never reaches a human being at all. Nobody made a decision, or so everyone involved will insist afterward.

Kalenok traces what happens next. The recruiter could argue the system only provided a recommendation. The vendor could argue the employer decided how to use it. The employer could argue a person made the final call. “All of those things can be true while the candidate with 61 never had a human look at the application,” she writes.

That circular defense is a liability sitting inside your recruitment stack right now. The party left holding it, when someone finally asks who decided, is not the vendor who built the model or the recruiter who trusted the score. It is talent acquisition.

Why the Accountability Gap Lands on TA, Not the Vendor

A vendor sells a scoring tool. Your organization chooses to deploy it, decides which roles it screens, and sets the threshold that separates an interview from a rejection nobody ever sees. When a regulator or a rejected candidate asks who decided, the honest answer inside most recruiting functions is nobody, precisely, and that answer will not hold up. The decision was distributed across a model’s training data, a vendor’s default settings and a recruiter’s unexamined trust in a number. None of those three will show up to account for it. Talent acquisition will.

Recruiters have always made bad calls too, rejecting the right candidate for the wrong reason. Kalenok concedes as much. The difference is that a human decision leaves a name attached to it, while an algorithmic one does not, unless someone inside the organization forces it to.

The rejection nobody can explain six months later is usually the one that gets challenged. By then, the evidence trail that would have answered the question does not exist. The gap is not that AI screens candidates. It is that most organizations using it cannot produce a record of why any single candidate was screened out.

The Adoption Figures Are LinkedIn’s Own Account

Kalenok cites LinkedIn’s 2025 Future of Recruiting research, which found that 37% of recruiting organizations were actively integrating or experimenting with generative AI. The figures come from LinkedIn’s own research, worth reading as a signal of direction rather than a verified industry baseline.

Organizations already using or experimenting with generative AI reported saving around 20% of their workweek, LinkedIn’s research found. Kalenok does not dispute that this productivity gain is real, or that summarizing resumes and drafting job descriptions is a legitimate use of the technology.

The problem starts when the same systems move from organizing information to scoring and ranking candidates. LinkedIn’s 37% figure measures general generative-AI adoption, not how many of those organizations use it specifically to score or rank candidates. The research does not draw that distinction. Kalenok makes the mechanism explicit: “A system can potentially repeat the same bad assumption across hundreds or thousands of applications before anyone realizes something is wrong.” Scale without an audit trail turns one bad assumption into a pattern nobody catches.

The High-Risk Label Already Exists

The EU AI Act does not treat this as hypothetical. It classifies certain AI systems used in employment as high-risk, placing recruitment alongside other automated systems capable of a meaningful impact on someone’s life. Employment law has spent decades defining boundaries around how humans make hiring decisions, and that history is exactly what a scoring algorithm complicates. There is no boundary left to enforce if nobody inside your organization can say who decided. If a regulator asked your team tomorrow to show why one candidate never made it past a score, could you answer?

Whether an algorithm is qualified to screen candidates matters less than who can show why it screened out the ones it did. In our view, if your organization operates anywhere the AI Act’s employment provisions reach, or in any jurisdiction building comparable transparency expectations, “we do not know why the system scored this candidate a 61” would look far closer to an admission than a compliance posture.

The Unconventional Candidate the Model Was Never Trained to See

Kalenok’s sharpest observation sits underneath most AI screening deployments: companies say they want transferable skills, unconventional backgrounds and different perspectives, while training systems to recognize what a successful candidate has historically looked like. A model built on years of past hires gets very good at finding more of the same. That pattern is exactly what makes the career pivot, the nontraditional degree, or the candidate who looks nothing like the role’s current incumbents harder to see.

An auditable rationale for every automated rejection does not just satisfy a regulator. It forces someone inside your organization to examine why that candidate scored low, on what basis, and whether that basis reflects the skill or just the resemblance. Without that record, the filtering happens silently, and the contradiction Kalenok describes gets worse every hiring cycle, not better.

The cost of skipping that record is real, not abstract: a rejected candidate who never gets an answer, and a talent acquisition leader who discovers, too late, that nobody decided was never going to be an acceptable answer.

Original reporting: Mexico Business News.