A recruiter opens a shortlist. The top candidate carries a 92% match score. The second sits at 89%. The recruiter looks at both profiles, cannot see what separates them, and does what recruiters have always done: ignores the ranking and reads the CVs.<\/p>\n\n\n
That moment is where most recruiting AI quietly fails. Not because the model was wrong, but because it gave an answer no one could act on.<\/p>\n\n\n
Key takeaways<\/p>\n\n\n
A number compresses a judgement into something that looks precise and communicates nothing. Ninety-two percent of what? Measured against which requirement? If the score drops to 74%, what changed?<\/p>\n\n\n
Recruiters are not resistant to being helped. They are resistant to being told a conclusion without the reasoning, because their job is to defend that conclusion to other people. A consultant who presents a candidate cannot say “the system ranked them first”. They have to say why.<\/p>\n\n\n
If a recruiter cannot repeat the reasoning in their own words, the tool has not saved them work. It has added a step they will skip.<\/p><\/blockquote>\n\n\n
Three conversations the score has to survive<\/h2>\n\n\n
Every match a system produces gets tested in at least three places, and a bare number fails all three.<\/p>\n\n\n
With the client.<\/strong> “Why this person and not the other four?” The answer has to reference the brief, this specific experience, that specific environment. A percentage invites the client to ask what the missing 8% was, which is not a conversation anyone wins.<\/p>\n\n\n
With the candidate.<\/strong> Approaching someone credibly means knowing why you are approaching them. Generic outreach built on an unexplained score is indistinguishable from spam, and experienced candidates recognise it instantly.<\/p>\n\n\n
With the hiring manager, six weeks later.<\/strong> When a placement struggles, someone will ask how this candidate was selected. “The tool suggested them” is not an answer that survives that meeting.<\/p>\n\n\n
What makes a match explainable<\/h2>\n\n\n
Three properties, and all three have to be present.<\/p>\n\n\n
\n
- Evidence.<\/strong> The match points to something concrete in the candidate\u2019s history, a role, a scope, a technology, a sector, not to a similarity score between two vectors.<\/li>\n\n
- Provenance.<\/strong> You can see where that evidence came from. A line in a CV, a note from a call two years ago, a field in the record. If the system cannot show its source, it is asserting, not reasoning.<\/li>\n\n
- Absence.<\/strong> The match says what is missing<\/em> as clearly as what is present. A candidate who fits on four of five criteria is useful information. A candidate presented as a 92% match, whose gap you discover in the client interview, is a damaged relationship.<\/li>\n<\/ul>\n\n\n
That third one is the least common and the most valuable. Systems are built to sell confidence, so they surface matches and hide gaps. Recruiters need the opposite: they need to know where the risk is before the client finds it.<\/p>\n\n\n
Explainability and compliance are the same requirement<\/h2>\n\n\n
This is usually treated as two separate topics, one for the product team, one for the legal team. In recruiting they are one topic wearing two hats.<\/p>\n\n\n
Regulators looking at automated tools in hiring keep arriving at the same demand: be able to describe how the tool works and how a decision was reached. In New York City, Local Law 144 obliges employers using automated employment decision tools to run independent bias audits and notify candidates. In Europe, the AI Act places recruitment and candidate selection among its high-risk categories, with transparency obligations attached. The specific requirements differ and will keep evolving, this is not legal advice, and anyone deploying these tools should take proper counsel.<\/p>\n\n\n
But the practical consequence is stable: a system that cannot explain its output cannot be audited, documented, or defended. The feature that makes a recruiter trust a shortlist is the same feature that makes the tool defensible. Build one and you get the other.<\/p>\n\n\n
What to ask before you buy<\/h2>\n\n\n
Four questions, and the answers are usually revealing.<\/p>\n\n\n
\n
- Show me a match and tell me why it ranked there, without using the word “score”.<\/li>\n\n
- Show me what this candidate is missing relative to the brief.<\/li>\n\n
- Where did that piece of evidence come from, and when was it recorded?<\/li>\n\n
- If I disagree with a ranking, what happens? Can I correct it, and does the system learn or ignore me?<\/li>\n<\/ul>\n\n\n
A vendor who answers the first two comfortably is building for recruiters. A vendor who deflects to model architecture is building for a demo.<\/p>\n\n\n
The point is not the ranking<\/h2>\n\n\n
The goal of matching is not to replace a recruiter\u2019s judgement with a number. It is to get a well-argued shortlist onto their desk faster than they could assemble it alone, with the reasoning attached so they can challenge it, correct it, and defend it.<\/p>\n\n\n
A ranking a recruiter cannot explain is a ranking they will not use, and a tool nobody uses solves nothing.<\/p>\n\n\n
That principle is why Twenty One Nexus<\/a> shows the reasoning behind every ranking rather than a confidence number.<\/p>\n\n\n\n