AI CV screening is easy to demo and hard to defend. We optimised from the first day for explainability and control rather than the cleverest answer, building on structured output, human-controlled autonomy and an audit trail that can answer for any decision months later.
Defensible before smart
It takes an afternoon to build a CV screener that feels magical in a demo: paste a job description, paste a CV, ask a model who is a good fit. It takes real design to build one a company can actually put in front of candidates. Automated hiring decisions are regulated, and increasingly so.
New York City's Local Law 144 mandates bias auditing and disclosure for automated employment decision tools, the EU AI Act classifies hiring AI as high-risk, and GDPR gives candidates rights around decisions made solely by automation. A screening feature that cannot explain what it did and why is a liability wearing the costume of a feature. So we optimised, from the start, for explainability and control rather than for the cleverest possible answer.
Structured, not freeform
The first decision was to never ask the model for an opinion in prose and then try to parse it. Freeform output is unpredictable in shape and impossible to audit consistently. Instead we require the model to return a structured result through a strict, tool-shaped schema: a numeric score from zero to one hundred, a short summary, a recommendation, and, the part that matters most, a verdict on each requirement, each one accompanied by a direct quote pulled from the CV as evidence.
Every judgement the system makes points back to specific text in the document it read. The output has a deterministic shape we can store, display and reason about, and the reasoning is inspectable rather than merely asserted. That is the difference between a score a recruiter has to take on faith and one they can check against the words on the page.
Autonomy is a dial, not a switch
The second decision was to refuse the binary of the AI does the screening versus it does not. We built four autonomy modes and let each organisation choose per job. Off means no screening at all. Score-and-recommend is purely advisory: the model produces its assessment and a human decides everything. Auto-shortlist lets the system advance strong matches on its own but never reject anyone. Full auto additionally allows it to reject weak matches, within thresholds.
The default posture is human-in-the-loop; full automation is an explicit, bounded opt-in, not something that creeps in. A team can start with advice and earn their way to more autonomy as they build trust in the results. We were careful not to build a system whose most powerful setting is also its most silent, so even full automation supports candidate-facing disclosure where it is required and surfaces a periodic spot-check digest so a person retains oversight.
Never act on a bad read
Real CVs are messy: scanned images, unusual PDF encodings, layouts that turn into word soup when you extract the text. A screener that treats a garbled extraction as a genuine weak candidate will quietly reject good people for a technical reason that has nothing to do with them. So we compute a parse-confidence score for every CV and gate the system's autonomy on it.
A low-confidence extraction is never auto-rejected; it is surfaced for a human to look at. The score bands that drive advancement and rejection are explicit, named thresholds rather than a vibe, and rejection carries a stricter confidence floor than advancement does. That asymmetry is intentional and, we would argue, ethical: wrongly rejecting a strong candidate is a worse error than wrongly advancing a weak one, so the system is built to be more cautious about the harm that cannot be undone.
An audit trail that holds up
The third non-negotiable was that every decision must be reconstructable. For each screening we store the model used, the prompt version, the exact requirements in force at the time, the score, the evidence and the autonomy mode that produced the outcome. Prompt versioning deserves special mention: without it, the model changed and we do not know how the old decisions were made is an unanswerable question, and unanswerable questions are precisely what auditors and regulators pursue.
With a versioned, evidenced record, we can say exactly what the system saw and why it acted, months later, for any candidate. Forcing structure, gating on confidence and recording everything costs some raw flexibility and adds discipline and storage we could have skipped. What it buys is a screening system a compliance-minded buyer can adopt, a hiring manager can trust and a candidate can be told about, because every decision is bounded by human-set autonomy, backed by evidence from the CV itself, and preserved in a record that can answer for it later.