How AI CV screening actually works, and where humans stay in charge
A look inside Recruit’s screening: structured extraction, job-aware ratings, and the line we won’t cross.
The Narbad Team
Building the suite
When people hear “AI screening” they usually imagine a black box rejecting candidates. That is not what we built, and we think it is worth explaining exactly what happens when a CV lands in Narbad Recruit.
Step one: extraction, not judgment
The first pass has nothing to do with scoring. We parse the document, PDF or DOCX, and extract a structured profile: work history with dates, education, skills, languages, and contact details. This alone fixes a real problem: candidates stop being filenames and become searchable, comparable records.
Extraction is also where fairness starts. A structured profile means the reviewer sees the same fields for every candidate, in the same order, regardless of how beautiful or plain the original document was.
Step two: a job-aware rating with visible reasoning
Next, the profile is evaluated against the actual job description, not a bag of keywords. The model produces a rating and, importantly, the reasoning behind it: which requirements look strong, which look weak, and what is worth probing in an interview.
Recruiters can open that reasoning on every candidate. If it is wrong, they override it, and the override is what counts. The rating is an ordering aid for the reviewer’s attention, not a verdict.
The line we won’t cross
Recruit never auto-rejects a candidate. There is no threshold below which people disappear unread. AI ranks, summarizes, and drafts; humans shortlist, interview, and decide. We think that split is not just safer, it produces better hires, because the tool’s job is to make sure attention goes where it matters, not to replace it.