AI resume screening: how it works and how to use it responsibly

What AI resume screening is

AI resume screening means using software to read incoming CVs and help you decide whom to look at first. The software extracts the text, compares it with the job's requirements, and returns an ordered list. Some tools add a short summary of each candidate.

The point is triage. A role with 150 applications cannot get 150 careful reads in an afternoon. Screening software gives you a defensible order to work through, so the effort goes to the people most likely to fit.

How it works

Tools differ, but most follow the same shape.

• Read: the file is converted to text. See the resume parser guide for why layout matters at this step.
• Define: the job's requirements are listed, usually split into must-haves and nice-to-haves.
• Match: the tool checks how much of the requirement list the CV covers.
• Add signals: some tools also look for context such as certifications, leadership or seniority.
• Rank: candidates are ordered by a score, ideally with the evidence shown.
• Review: a person reads the top of the list and decides.

What it can and cannot do

A CV that does not mention a skill scores as if the candidate lacks it, even when they have it. That is why the list is a reading order, not a verdict. Screening also says nothing about people who never applied, so a good process pairs it with outreach and with a clear, honest job description. The JD Enhancer in Rezumz can tighten a draft before you open the role, which helps because the clearer the requirements, the cleaner the ranking.

How Rezumz ranks candidates

Rezumz scores each CV against the job's must-have and nice-to-have skills. The score is calculated from the skills found in the CV text. 80 percent is the share of must-haves found and 20 percent is the share of nice-to-haves found, allowing for slightly different wording. Missing a nice-to-have barely changes the score, while missing a must-have changes it a lot.

The AI also reads each uploaded CV for extra signals, such as certifications and leadership, and shows them next to the score. It adds a short summary of strengths. Applications through the apply page are scored without AI.

Three design choices matter for fairness. The score is deterministic, so the same CV against the same job gives the same number. Nobody is rejected automatically: the score orders your list and never rejects, emails or moves anyone. And you can see why, because the lines of CV text behind a match are shown. If scores look wrong, the usual cause is the skills list, so correct the must-haves and add the CVs again.

Bias and legal cautions

Software can repeat bias in the requirements you give it. A must-have such as a specific degree, or an exact number of years, can quietly exclude capable people. Review must-haves critically before you run anything.

Rules about automated or partly automated hiring decisions differ by country and, in some places, by city or state. Some places give candidates rights to be told about automated processing or to ask for human review. This is not legal advice. Check your local rules or take advice before relying on any tool.

• Keep a human reviewer on every decision that affects a candidate.
• Keep an audit trail of what was scored, what was decided and why.
• Do not use signals that proxy for protected characteristics, such as photos, age or names.
• Record a reason when you reject someone, so patterns can be reviewed later.
• Tell candidates how their data is processed.

A screening workflow for 150 applications

This workflow is a suggestion, and any times are examples rather than measured results. It assumes one recruiter and a hiring manager who can review a shortlist within two days. Screening is one early step: after it come shortlisting, interviews, structured feedback and an offer, so keep your notes where the rest of the hire lives.

Marketing pages for screening tools often say little about how a ranking is produced, so ask vendors direct questions. Is the score a fixed calculation or a model output that can change between runs? Can you see the text matched for each requirement? What happens to a CV it cannot read? Which decisions happen without a person? For hiring, the right answer to the last one is none.

• Step 1, fix the criteria: from the brief, write 4 to 6 must-haves and 3 to 5 nice-to-haves. Remove any that are preferences, not requirements.
• Step 2, load all 150 CVs through your apply page, upload or inbox, and check that none came back empty. Empty results are often scans.
• Step 3, read the top 30 by score in full, using the quoted evidence to speed up your reading.
• Step 4, spot-check 10 from the middle and 5 from the bottom to catch people the wording missed.
• Step 5, shortlist 8 to 12 and note why each is in.
• Step 6, record a reason for each rejection, and send respectful replies.
• Step 7, review the shortlist with the hiring manager before scheduling interviews.

Checklist to evaluate any AI screening tool

• Can you see why a candidate ranked where they did, with the supporting text?
• Is any decision automatic, such as rejections or emails sent without a person?
• Does the same CV give the same score each time?
• What happens to a file it cannot read: an error, or a silent zero?
• Where is data stored, how long, and is it used to train models?
• Can you edit the criteria, and re-run after editing?
• Is there an audit trail?
• Can you delete a candidate on request?
• What does it cost at your volume? Check the vendor's current pricing.

What good criteria look like

Screening is only as sound as the list it checks. A good must-have is specific, testable from a CV, and truly required on day one. A weak one is a preference or a proxy.

Checking the tool for yourself

Before a real vacancy, run a dry test. Take a role you filled last year. Load ten CVs, including the person you hired, two you rejected for good reason, and a few you thought were close. See whether the ranking puts the hire near the top and whether the evidence shown makes sense. If a strong person scores low, find out why: was it the skills list, a parsing problem, or wording the tool did not match?

Run the test again after you change the criteria, so you know the effect of each change. Keep the results. They are useful if anyone asks how you chose a tool. Also try the awkward cases: a career changer, a school leaver, someone with a gap, and someone whose CV uses different words for the same skill. Those are the people a rigid matcher is most likely to undervalue.

Common mistakes

• Treating the score as a verdict and skipping the read.
• Setting too many must-haves, which shrinks the pool to almost nobody.
• Copying a job description as the criteria, including its wish list.
• Never testing on CVs you already know the answer for.
• Ignoring files that came back empty.
• Skipping the audit trail because the list looked obvious.

Can I use ChatGPT to screen resumes?

Honestly, yes, up to a point. A general chat assistant can summarise a CV, compare it with a job description, and draft screening questions. For ten applications it can be useful.

It breaks down at volume and under scrutiny. Answers can vary between runs, so two identical CVs may be ranked differently. Pasting candidate data into a chat tool raises privacy questions about storage and training, so read the provider's terms and your own policy first. You must paste CVs in one at a time or in batches, and long batches can lose detail. There is also no audit trail, no fixed scoring rule, and no record of why someone was ranked lower.

A dedicated tool is better when you screen regularly, have more than a few dozen applicants, need consistent scoring, or must show how decisions were made. Rezumz scores with a fixed rule, shows the evidence, and keeps a record. If you only need a one-off summary, a chat assistant may be enough.

Who it is not for

Rezumz suits small and mid-size hiring teams. If you need deep HRIS or payroll integration, or OCR for scanned CVs, check whether it covers your case in the demo. The free plan includes 2 free candidate analyses, the JD Enhancer and the full scoring breakdown, so you can test it on real applicants first.