AI resume screening: what a tool may prepare and what a human must decide

A tool may prepare a hiring decision; it may not be the decision, because a rejection signed by nobody is abdication wearing automation as a costume.

Every hiring round begins the same way: a posting goes up, and hundreds of applications arrive, most from people who will never be right for the role. Somebody has to read them, and reading them is genuinely hard labor. So the tool arrives with its promise: it will read everything, rank everyone, and hand you a short list. That promise is real, and it is also where the trouble starts, because a tool that sorts people learns from the past, and it inherits every pattern the past contained, including the ones nobody intended.

Screening can be honest labor. Reading hundreds of applications against the same fixed questions, the same way, every time, is a kind of fairness a tired human at midnight cannot match; the tool applied the same bar to everyone who applied. Screening can also be abdication. When the reject letter goes out signed by nobody, written by nothing, decided by a score the company never examined, the company has not made a decision; it has laundered one. The line between the two is the whole subject: a tool may prepare a human decision, a tool may not be the decision.

The loop deserves its own warning, because it closes quietly. Hire only whom the tool ranked high, and next year's hires become next year's examples: the new benchmark of what a good application looks like. Feed those examples back in, and the tool teaches itself its own taste, round after round, until the company honestly believes the pattern is the market. Nobody chose to narrow the door. The system did it one efficient round at a time, and every step felt like diligence.

The practice starts with a name. Every rejection should have a human who owns it: not the person who read the short list, the person whose decision it was. Write that name down before the tool runs, because a decision that belongs to nobody in advance will belong to the tool by default. When a candidate asks why, an accountable human can answer; an accountable human can also notice when the ranking says something the company does not actually believe.

Then interrogate the vendor, and notice that the question is not about accuracy. Ask what the tool optimizes for: what outcome its score rewards, what it treats as a good hire, what its training data was. An honest vendor can answer in a sentence, and the answer is rarely the best employee; it is usually the candidate who resembles past hires, which is a different thing wearing the same words. A vendor who cannot answer plainly has told you something more important than the feature list.

Two habits keep the human in the loop where it counts. Tell candidates a tool is in the process; people forgive a system far more readily than they forgive a discovery, and disclosure is cheaper than the apology. And keep a path for the atypical CV: the career that changed fields, the self-taught engineer, the gap with a reason. Those are exactly the applications a pattern cannot score, and they are where a small company finds people the market sorted out.

We sit on the structure side of this: we build and advise on AI that says what it is, shows its grounding, and hands the consequential judgment to a person, and hiring is one of the places we draw that line out loud. What the law requires about automated decisions about people, and how applicant data may be kept and used, belongs to counsel and to the data rules of each jurisdiction, not to a vendor's marketing page. Guidance, not guarantees: the tool can carry the reading; the judgment stays human.