AI in recruiting: faster screening without screening out great people

We use AI in our own recruiting funnel every day, so this is not a sceptic's take. It is a practitioner's warning: the same tooling that made our shortlists faster can, configured carelessly, make shortlists worse — at scale, silently, with a confident ranking score attached.

What AI is genuinely good at in hiring

Parsing unstructured CVs into comparable profiles. Matching stated skills against role requirements, including the synonyms and adjacencies a keyword search misses — the SAP FICO consultant whose CV says "S/4HANA Finance," the data engineer who writes "Delta Lake" instead of "Databricks." Ranking a pool of three hundred so a human reads the most promising thirty first. Drafting outreach that reflects the actual role instead of a template.

All of this compresses time. Our own requisition-to-shortlist window dropped from weeks to about two days when the stack matured, and shortlist quality — measured by interview-to-offer conversion — went up, not down.

Where it quietly goes wrong

The failure modes are rarely dramatic. They're statistical, and they hide inside reasonable-sounding scores.

• Proxy bias. A model trained on "profiles like our past hires" learns your past, including the parts you'd rather not repeat — schools, employers, gaps, even name patterns.
• Format punishment. Great candidates with unconventional CVs — career changers, returners, people from non-corporate backgrounds — parse badly and rank low for reasons that have nothing to do with ability.
• Keyword theatre. Once candidates learn a machine reads first, CVs optimise for the machine. Ranking rewards the best CV-writers, not the best engineers.
• Confidence laundering. A ranked list feels objective. Recruiters defer to it even when their own judgement disagrees — the score becomes the decision instead of an input.

The five rules we run

1. AI ranks; humans reject. No candidate exits the funnel on a model's score alone. A person reads before anyone is out.
2. Score the requirement, not the résumé style. Match against explicit role criteria, never "similarity to previous hires."
3. Audit the tails monthly. Sample profiles the model ranked low and have a senior recruiter blind-review them. Every miss teaches you where the model is blind.
4. Keep the interview human and technical. The model gets people to the door faster; a practitioner still opens it.
5. Tell candidates. Transparency about AI-assisted screening costs nothing and increasingly, regulation requires it.

AI in recruiting is not optional anymore — the funnels are too large and the good candidates gone too fast. But the firms that win with it will be the ones that automated the reading, and refused to automate the judging.