Using AI to screen without losing the candidates you actually want
Automated screening now sits in most hiring funnels. The teams getting value from it are the ones who treat it as a ranking tool, not a rejection tool.
Ananya Rao
Director, Technology Practice · 8 September 2026 · 6 min read
Nearly every mid-size employer we work with has introduced some form of automated screening in the last two years. The promise is obvious: a recruiter reviewing 800 applications for a single engineering role cannot give each one a fair reading, and the ones at the bottom of the pile are rejected by fatigue rather than by judgement.
What we see in practice is more mixed. Screening models trained on a company's own historical hires tend to reproduce that company's historical blind spots. If your last twenty backend hires all came from three universities, a model that learns from those hires will quietly rank a fourth university lower — not because the candidates are weaker, but because the training data never gave them a chance.
The employers getting genuine value share one habit: they use the model to order the queue, never to empty it. A human still reads the top of every band, including the bottom band, on a sample basis. That sampling is what catches drift, and it costs a fraction of the time the ranking saves.
The second habit is measuring the right thing. Time-to-screen is easy to improve and tells you very little. The number that matters is offer-acceptance among candidates the model ranked in the bottom half — if that number is healthy, your model is discarding people you would have hired.
Our own matching engine is built on this principle. It surfaces and ranks; consultants decide. We publish the features it uses to every client, and we audit shortlist composition quarterly. If a model cannot be explained to the person being assessed by it, it should not be making the assessment.