AI in HR: How to Screen Candidates (and How Not To)
Using AI in HR: candidate screening, fit scores, bias risk, human oversight and candidate transparency.
A hundred applications arrive for one vacancy. Reading them all takes days, and the right candidate may sit among the ones nobody reached. AI opens that bottleneck — but creates a new risk where it opens: fast decisions you cannot defend.
Note: this article is about process design, not legal advice. Consult your legal adviser about your obligations in recruitment.
What AI is genuinely good at: ranking
Models are weak at splitting candidates into “suitable / unsuitable”. They are good at ranking by closeness to the role and summarising why. The difference looks small and changes the process completely:
Right use
A hundred applications ranked with reasons. The recruiter reads the top 25, while the rest are queued rather than rejected. The decision stays human.
Wrong use
The model auto-rejects 60 applications and nobody looks. The failure is silent: a rejected good candidate never comes back, and the mistake is never noticed.
What it can and cannot assess
- Can assess: evidence of required technical skills, length of experience, sector proximity, languages, location fit.
- Be careful with: vaguely defined criteria such as “culture fit”. A model cannot measure that; it guesses, and the guess carries bias.
- Should not assess: age, gender, marital status, origin, photograph. These should never enter the evaluation — if they do, discrimination risk follows.
- Never: the final rejection decision. The model ranks; a person decides.
The bias risk is real
Models learn from past data, and past hiring data carries past bias. A setup that picks candidates “resembling our successful employees” systematically filters out difference. To reduce it:
- Keep the criteria written and explicit. The model should be able to say what it looked at.
- Mask identity information. Name, photograph and age can be removed from the evaluation input.
- Audit the results regularly. If particular groups consistently rank at the bottom, the criteria are wrong.
- Make human review mandatory. No candidate should be rejected on a score alone.
Transparency towards candidates
Telling candidates that AI is used in the process is both right and increasingly expected. The reason for rejection also has to be one a person can give: “the system did not consider you suitable” is not a reason, and it ends the relationship — a candidate who is wrong today may be exactly right in two years.
A flow that works
- The vacancy is published and applications collect in one pool.
- The model scores each application against the role and writes its reasoning.
- The recruiter reviews the ranked list with reasons, and can also see low scores with weak reasoning.
- A person selects who is invited to interview.
- Rejections go out with human approval.
- Suitable-but-not-now candidates stay in the pool — the best list for the next vacancy.
Measurement
- Screening time: days from application to first interview. The clearest gain shows here.
- Score-to-outcome alignment: were high scorers actually hired more often? Poor alignment means the criteria are wrong.
- Human override rate: how often does the recruiter break the ranking? Very high means the model is not working; zero means nobody is looking.
- Candidate experience: response times and the share who received an update.
Frequently asked questions
Can the model spot exaggeration in a CV?
To a limited degree. It can notice inconsistencies, such as titles that conflict with the length of experience, but verification belongs to the interview. Screening is a ranking, not a filter.
Is it needed for small teams?
Not at low application volume — reading ten applications is already fast. The benefit appears as volume grows.
How long should candidate data be kept?
Limited to the purpose and for a defined period. If you want to keep people in a pool, you need to tell them and set a retention period.
Next step: to see the candidate evaluation flow with one of your own vacancies, request a demo.