The line AI shouldn't cross in hiring

If you're considering bringing AI into your hiring process, you've probably heard two opposite messages at once: that adopting it is essential, something your HR team can't afford to fall behind on, or that it's a shift you can't fully control, one that exposes the company to legal and reputational risk. The truth, in my experience, sits somewhere in between, but it's fairly clear exactly where the line falls.

This piece is about finding that line, and about what happens when it gets ignored.

Where AI genuinely helps

Industry data backs up what many recruiters already sense intuitively: AI's biggest real benefit shows up at both ends of the process. Before it even starts, that means drafting job postings and pulling scattered requirements from hiring managers into a coherent brief. Once the process is underway, the same help shows up during the interview itself and right after it: automatic transcription, notes generated on the fly, and interview summaries free the recruiter from the need to run a conversation and write everything down at the same time. That means actually listening to the candidate, keeping eye contact, reacting to what they say, instead of splitting attention between the conversation and the keyboard, and only afterward returning to a solid, ready-made summary instead of reconstructing it from memory. The same applies to the paperwork around the rest of the process: reminders, summaries for the team, prep for the next stage.

The effect shows up in the numbers. According to SHRM's 2025 Talent Trends report, 89% of organizations using AI in recruiting report time savings or increased efficiency. That's consistent with what we see in practice: transcribing conversations, generating notes automatically, or planning next steps genuinely takes work off HR teams' plates, work that never required human judgment to begin with, just patience and time.

Importantly, this doesn't mean the recruiter disappears from the process. It means a shift in what they focus on. The time recovered through automating logistics goes back to where it's actually needed: the conversation, the assessment of fit, the relationship with the candidate.

But a decision is a completely different category

This is where it's worth being genuinely cautious with the enthusiasm. Automating notes is one thing. Letting an algorithm decide who gets invited to a conversation, or who gets an offer, is an entirely different category of risk, and the data shows that risk is real, not theoretical.

The best-known example is Amazon's internal system, built starting in 2014 to automatically score résumés on a one-to-five-star scale. The model was trained on ten years of applications the company had received, and because tech was a male-dominated industry, most of those résumés came from men, so the system taught itself to prefer male candidates, downgrading résumés that contained the word "women's" (as in "captain of the women's chess club") or the names of certain all-women's colleges. Even after engineers manually patched those specific signals, the company couldn't be confident the model hadn't found other, hidden ways of reproducing the same pattern, and it ultimately scrapped the system. This illustrates a mechanism that should worry anyone thinking about deploying AI at scale in hiring: a model trained on data that reflects historical inequities reproduces them with total consistency, and the scale of the problem often only becomes visible once someone starts looking for patterns in the outcomes rather than in the code.

Another prominent, and still open, example is Mobley v. Workday. In June 2026, a federal judge ruled that discrimination claims against Workday's applicant-screening system could proceed to trial. The case was brought by a candidate who, after being rejected from more than a hundred applications, often within minutes, at all hours of the day and night, concluded that his applications most likely never reached a human being. Workday has consistently maintained that its systems don't make hiring decisions and that customers retain full control over the process (an important caveat), but the mere existence of the case shows how easily the line between "supporting the process" and "making the decision" blurs in practice.

Why the human eye is still irreplaceable

These two failures don't share the same shape. Amazon caught the problem itself: internal testing surfaced the bias, engineers tried to patch it, and the company ultimately killed the tool rather than risk shipping it. The Workday case is still unresolved and disputed, but the core question driving it is different: not what the algorithm learned, but whether anyone was actually reviewing its outcomes case by case, or whether a rule, once trained, simply ran unchecked at scale for years. What both cases point to is the same underlying risk: an automated system can apply a pattern to thousands of candidates with a speed and consistency no single recruiter could match, and that scale is exactly what makes it dangerous when nobody is checking the individual outcomes.

That's exactly why the final judgment call (whether this person fits the team, whether a gap in the résumé has a reasonable explanation, whether an unconventional career path is an asset or a problem) has to stay with a human being. No system, however well trained, can weigh context the way a person sitting across the table can. It can flag a pattern, but it can't ask a follow-up question, notice a shift in tone, or decide that a rule doesn't apply to this particular person for a reason no dataset would ever capture.

The practical line

If I had to boil this down to one rule that would serve any team rolling out AI in recruitment, it would be this: automate everything that's logistics, not judgment. Scheduling, transcription, summaries, reminders, initial sorting of documents against formal requirements: that's a safe, high-payoff zone for automation. Shortlisting candidates, ranking them: that's useful advisory input, something a recruiter still reviews and can override. The final call on who gets invited or hired is different: that's the one thing that has to stay with a person, never an algorithm.

None of this comes from distrust of the technology. AI is valuable, worth keeping up with, and genuinely speeds up a lot of what used to slow recruitment down. But like any of us, it can get things wrong, or quietly repeat bad patterns, if no one's watching. That's exactly why the decision needs to stay with a human, probably for a long time yet.