Screening is the part of AI recruiting that makes people the most nervous — and reasonably so. Ranking a shortlist is one thing; deciding who advances and who doesn't touches someone's livelihood. The teams that get real value out of AI screening are the ones who are precise about which parts of the decision they're willing to automate, and which parts they aren't.
What models are actually good at
Pattern matching across a candidate's documented work — tools used, years in a domain, career trajectory, overlap between a job description and demonstrated experience — is exactly what these models are built for. It's fast, consistent, and free of the fatigue that creeps into a recruiter's fiftieth resume review of the day.
Where it should stop
- Anything that functions as a proxy for a protected characteristic, even indirectly.
- Final hire/no-hire decisions — screening should narrow the field, not make the call.
- Judgment about culture fit, communication style, or anything that only shows up in a conversation.
“A good screening model tells you who's worth a conversation. It should never pretend to have already had one.”
In practice, this means every AI evaluation we ship comes with its reasoning attached — the specific experience, skills, or signals that produced the score — so a recruiter can check the model's work in seconds instead of trusting it blindly. If you can't see why a candidate was ranked the way they were, the ranking isn't something you should be acting on.
The goal isn't a model that never needs oversight. It's one that makes oversight fast enough that you actually do it, every time, instead of rubber-stamping a black box because checking would take too long.