AI résumé screening: how it works, from PDF upload to ranked shortlist

What AI résumé screening is, and what it is not

AI résumé screening (also called AI CV screening) is software that reads the applications for a role and helps a recruiter decide which ones to look at first. It takes over the slowest part of the first pass: opening every PDF, hunting for the same facts, and comparing them with a mental checklist of what the role needs.

It is not the same as an applicant tracking system (ATS), which manages the hiring pipeline: job posts, application stages, interview scheduling and offers. Many ATS products include keyword filters, but a keyword filter only checks whether a word appears. AI screening tries to answer a different question: does this candidate meet this requirement, and what in the CV shows it?

It is also not a hiring decision. The output is a ranked list with reasons, and its value depends on whether a person can check those reasons quickly. Teams that treat the top score as an automatic yes, or a low score as an automatic no, turn a review tool into an unreviewed filter, with the risks described in our guide to fair AI hiring and human review.

How AI résumé screening works, from PDF to shortlist

A résumé screening flow has five stages, and each one can be designed well or badly.

1. Define the role. The hiring team writes the requirements: skills, years and type of experience, qualifications, languages, and which of these are must-haves. This is the brief the AI will apply to every candidate.
2. Upload the CVs. Recruiters add PDF files to the role, one at a time or in batches. Each file is stored against the role so every result can be traced back to the original document.
3. Extract structured information. The system turns each PDF into fields: work history with employers, titles and dates, education, skills, languages and certifications. This is where layout and language problems appear.
4. Compare with each requirement. A language model, guided by the role definition, checks the structured profile against each requirement and writes a short judgement for each one.
5. Score, rank and explain. Each candidate receives an overall score and a written summary, and the list is sorted so the strongest matches appear first. The recruiter opens a candidate, reads the reasons and decides who moves on.

HireDash, the AI résumé-screening product FLUIDS designed and built, follows this shape. The hiring team defines requirements and scoring priorities, or uses a job description to draft them with the AI role builder, and the draft stays editable before any CV is screened. The team then uploads PDF résumés to the role, and each candidate is shown with a fit score from 0 to 100, a written summary and a breakdown against the role's requirements.

Parsing Arabic and English PDF résumés

Text extraction is the first place screening goes wrong, and errors here are invisible later: a model cannot match experience it never received. In Saudi Arabia and the wider GCC, hiring teams commonly receive CVs in English, in Arabic, and in both languages on the same page, which adds problems an English-only tool never meets.

• Scanned or image-only PDFs. Some CVs are photos or scans with no text layer. They need optical character recognition (OCR), which reads images as text and makes mistakes, more often with Arabic script.
• Two-column and designed layouts. Sidebars, tables and icons can make extracted text come out in the wrong order, so a job title ends up attached to the wrong employer or dates.
• Right-to-left text. Arabic extracted from some PDFs comes out with reversed word order or broken letter joins. Mixed lines, such as an Arabic job title next to an English company name, are especially fragile.
• Names and transliteration. The same name can appear as محمد, Mohammed, Muhammad or Mohamed. That matters for finding duplicates, and it is one more reason to keep names out of scoring altogether.
• Dates and digits. CVs may use Arabic-Indic digits, Hijri years, or open ranges such as 'present' and 'حتى الآن'. Experience calculations must normalise them before comparing with a requirement such as five years in a similar role.

The practical rule: when extraction is uncertain, say so. A tool should flag a CV it could not read reliably and route it to a person, rather than produce a low score that looks like a weak candidate. Before launch, test extraction on a set of real layouts in both languages, using synthetic CVs or documents whose owners have agreed to that use, and check the structured output field by field.

Requirement matching vs keyword matching

Requirement matching checks each requirement separately and records whether the CV shows evidence for it; keyword matching only counts words. The difference shows up in the first week of real use.

Custom questions extend the role definition. In HireDash, a recruiter can add custom extraction questions that are applied to every candidate, with answer types such as yes/no, a number, a category or a short summary. A question like 'Has the candidate managed a team?' (yes/no) or 'How many years in a client-facing role?' (a number) turns a soft requirement into something explicit and comparable across the whole list.

Write requirements the way an interviewer would test them. 'Strong communication skills' cannot be judged fairly from a CV; 'has written client-facing reports in Arabic and English' can be checked for evidence. The more testable each requirement is, the easier it becomes to tell a good ranking from a lucky one.

How to read and use a ranked shortlist

A ranked shortlist is an order in which to review candidates, not a verdict on who is good. Treat it the way you would treat a well-prepared first-pass summary from a colleague: useful, fast, and worth checking.

• Read the reasons, not only the number. In HireDash, the fit score sits next to written reasoning and a requirement-level breakdown, so a recruiter can see why a candidate ranked where they did.
• Treat close scores as a group. Small differences between scores rarely mean a real difference in fit. Candidates a few points apart deserve the same attention.
• Look below the cut-off. Review a sample of lower-ranked CVs for each role, especially in the first weeks. It is the quickest way to find a requirement written too narrowly or a CV that was parsed badly.
• Keep the decision with people. Who is invited to interview, and who is hired, is decided by the hiring team.

The 0 to 100 score in HireDash is a product field that summarises the comparison with that role. It is not a measured accuracy statistic, and it should not be presented to anyone as a probability that the candidate will succeed in the job. Our guide to explainable candidate scoring covers how scores and reasons should be designed so a reviewer can check them.

What to plan before building or buying an AI screening tool

The first decision is where screening sits in your process. If you already run an ATS, screening may be a step that reads its applications; if CVs arrive by email and shared folders, a standalone tool may come first. Our comparison of an ATS and a custom AI screening tool works through that choice. Either way, a first release should cover:

• Role definition with must-have and nice-to-have requirements, editable by the hiring manager.
• Reliable PDF extraction in the languages your applicants use, with uncertain files flagged for a person.
• Per-requirement reasons alongside every score.
• A record of which role definition and model version produced each result, so a ranking can be explained later.
• Data handling agreed in advance: who can see CVs, how long they are kept, and whether past applicants may be considered for future roles, as covered in reusing your talent pool.

FLUIDS designs and builds AI products and internal tools for companies in Saudi Arabia and the GCC, in Arabic and English. See the HireDash case study for the screening workflow, read how we approach AI features inside a product, or explore SaaS product development if you plan to offer screening to other companies.