Can AI screen résumés accurately?

It can read a CV against a brief reliably and cite what it found, which is the mechanical part and is worth automating. What it cannot do reliably is judge what a gap means about a person. The practical answer is to let software do the reading and comparison, and keep the decision with a recruiter who sees the evidence behind every conclusion.

Why do good candidates get rejected by screening software?

Usually because the tool treats a requirement the CV never mentions the same as one the CV shows is absent. A strong candidate with a short, factual CV scores badly against a keyword filter despite having the experience. Keeping 'not demonstrated' and 'not addressed' as separate states is what prevents it.

Is a match percentage useful at all?

Only when you can see how it was produced and what it excludes. A score computed from stated weights over cited evidence, shown next to the ceiling it would reach if the open questions resolved, tells you something. A score from a model that cannot explain itself is a number nobody can check, attached to a decision about a person.

Does this require replacing our ATS?

No, and it usually should not. Screening reads requisitions and CVs and writes back a shortlist with its reasoning. Building around an existing ATS is faster, cheaper and far less disruptive than a migration, and it keeps your candidate data where your team already expects to find it.

Guide

Screening CVs against a brief without a black-box score

Most screening tools answer the wrong question. They tell you how well a CV matches a brief, when what a recruiter needs to know is what the CV actually evidences and what it simply does not say.

Who this is for: Recruiters, staffing firm owners and talent teams evaluating screening automation, or wondering why the last tool did not survive contact with the desk.

9 min read · Last reviewed 25 September 2026

A recruiter with sixty applications and an afternoon does the same thing every time: skims for keywords, keeps the ones that look close, and moves on. That is not carelessness, it is arithmetic. But the cost is specific and it is always paid by the same people — a strong engineer with a terse one-page CV loses to a weaker one who wrote more.

Screening software was supposed to fix that. Mostly it automated it. A keyword filter is the same skim performed faster, and a model that returns an unexplained percentage has added confidence without adding information.

The distinction that matters most, and the one most tools erase

There are two completely different reasons a requirement is not ticked on a CV, and almost every screening tool treats them identically:

Status	What it means	What to do
Not demonstrated	The CV covers this ground and the requirement is genuinely absent	Evidence of absence — a real gap, worth weighing
Not addressed	The CV is silent; there is no evidence either way	Absence of evidence — a question to ask, not a mark against them

Collapsing those two is how a tool rejects someone for a skill they have and simply did not write down. It is worth being blunt about who that penalises: people who have not had to job-hunt recently, people writing in a second language, people who were told a CV should fit one page, and people without access to the kind of coaching that teaches you to mirror the job advert.

A tool that keeps the two apart behaves differently in a way you can see. The silent requirement produces a question for the screening call rather than a deduction, and the candidate stays in play until somebody has actually asked.

What is wrong with the percentage

An unexplained match score is worse than no score, because it is acted on. Once a number sits next to a name, it gets sorted by, filtered on, and eventually trusted — and nobody can check it, including the person who bought the tool.

Two things fix this, and neither requires giving up on scoring:

Show the arithmetic. Which requirements were weighted how, what each one scored, and what that sums to. A score anyone can recompute by hand is a score that can be argued with, which is the point.
Show the ceiling alongside the score. If a third of the brief is simply unaddressed by the CV, the score is a floor rather than an estimate, and presenting it alone understates everyone whose CV is thin.

A candidate at 46% whose CV never mentions half the stack, and a candidate at 37% whose CV names its whole stack and does not contain what you need, are two completely different situations. One is unmeasured. The other is measured and short. A single percentage cannot tell you which is which, and the difference decides whether a phone call is worth making.

See it working

We built a worked example of exactly this: four synthetic CVs against one brief, every requirement showing the line of the CV it was judged on, and the two candidates above sitting nine points apart for opposite reasons. Synthetic data, real logic, and nothing auto-rejects.

See the two side by side

Automate the reading, not the decision

Software is genuinely good at reading every CV properly, against every requirement, without getting tired at application forty. That is real work and it is worth automating.

What it is bad at is the thing it is usually sold for: turning that reading into a verdict on a person. A gap means different things depending on why it is there, and weighing that is a judgement with consequences for somebody's livelihood.

Extract the requirements from the brief and get them agreed. If the tool is matching against the wrong list, everything downstream is confidently wrong.
Read each CV against each requirement and cite the evidence — the actual line, not a similarity number.
Mark what is not addressed as a question, and generate the question in language a recruiter can ask as written.
Hand the recruiter a shorter job than they started with. Do not hand them a decision.

There should be no threshold at which a tool rejects anyone automatically. If a candidate falls below a bar, the useful output is the reason and the evidence, routed to a person — not a silent removal from the pile. Automated rejection is also the part most likely to attract legal attention, and the part hardest to defend after the fact.

Keep the record, because you will be asked

Screening is the one automation in a staffing business where somebody may later ask why a particular person was not put forward. The answer should be a record rather than a recollection: what the system read, what it concluded, what evidence it cited, and what the recruiter decided afterwards.

That audit trail is worth building on day one. Retrofitting it means reconstructing decisions from memory, which is exactly the situation it exists to prevent.

Common questions

Can AI screen résumés accurately?: It can read a CV against a brief reliably and cite what it found, which is the mechanical part and is worth automating. What it cannot do reliably is judge what a gap means about a person. The practical answer is to let software do the reading and comparison, and keep the decision with a recruiter who sees the evidence behind every conclusion.
Why do good candidates get rejected by screening software?: Usually because the tool treats a requirement the CV never mentions the same as one the CV shows is absent. A strong candidate with a short, factual CV scores badly against a keyword filter despite having the experience. Keeping 'not demonstrated' and 'not addressed' as separate states is what prevents it.
Is a match percentage useful at all?: Only when you can see how it was produced and what it excludes. A score computed from stated weights over cited evidence, shown next to the ceiling it would reach if the open questions resolved, tells you something. A score from a model that cannot explain itself is a number nobody can check, attached to a decision about a person.
Does this require replacing our ATS?: No, and it usually should not. Screening reads requisitions and CVs and writes back a shortlist with its reasoning. Building around an existing ATS is faster, cheaper and far less disruptive than a migration, and it keeps your candidate data where your team already expects to find it.

Related guides

Where a recruitment desk actually loses its week

Ask a recruiter where their time goes and they will say sourcing. Watch the desk for a week and it is rarely sourcing. It is the packaging, chasing and re-keying around it.

Three-way matching, and why it breaks in small manufacturing

Matching the purchase order to the goods receipt to the invoice is the oldest control in payables. It is also the one most small factories quietly abandon, and for understandable reasons.

If this describes a process you recognise — we are a small company in Vadodara that builds exactly this kind of thing. No obligation and no sales sequence: get in touch or try the automation finder, which will tell you when the answer is not to automate.

+91 93131 52136 · hello@twishivplatforms.com