What Actually Happens to a CV When AI Screens It: The Evidence Trail SMB Hiring Teams Can Inspect

Before a recruiter sees a single line, one CV becomes five machine-readable artifacts that create a paper trail: raw parsed text, extracted fields, knockout-question answers, ranking signals, and a recruiter-facing summary card. We inspect the outputs that decide a CV's fate at each stage—a critical step for bilingual Arabic/English CVs, where English-centric screening loses meaning fast.

Your team posts a role in the morning and gets 40 applicants by lunch. The system is meant to sort them. The real question is what it creates, hides, or rejects before anyone on your team even opens the file.

A CV arrives: the system starts making copies of meaning

The screening system never touches the original file.

The system instead creates several internal versions to work with:

• The original file. Untouched.
• A raw parsed text version.
• Structured fields for the system to query and rank.

That raw text version matters more than the PDF design. If the parser scrambles the text, every step downstream will work with broken data, compounding the initial damage.

You can see this directly in public parser documentation. Affinda’s resume-parser docs show a response with both rawText and structured fields like candidateName, email, and workExperience in the same object link. DaXtra’s sample output shows normalized fields such as skillName, skillUsed, and skillProficiency link.

As its opening move, the AI screening pipeline parses the resume and extracts structured data before making any judgments.

Stage 1: What the system extracts first — and what can disappear

The system extracts specific fields first. Workable’s AI Screening Assistant uses the candidate’s Profile Summary, Education, Experience, Skills, and attached Resume to build short summaries of candidate profiles for recruiters link.

The first pass typically extracts information such as:

• name and contact details
• job titles
• employers
• dates
• education
• skills
• location
• free-text summary content

The main challenge is that document parsers rebuild text based on a logical reading order instead of the visual layout. Microsoft’s Document Intelligence splits multi-column content into separate lines link. Amazon Textract reads columns top to bottom, left to right, treating tables and text as separate blocks link. Adobe’s PDF Extract API explicitly leaves headers and footers out of the output link. Google’s parser also calls out headers, footers, and tables, warning that tables spanning pages may be split link.

This process explains why certain formatting choices break hiring workflows, including:

• tables used for layout
• two-column CVs
• headers or footers holding contact info
• image-based design
• PDFs that look clean to humans but messy to machines

Using bilingual CVs introduces another potential failure point. Greenhouse documents full parsing capabilities in Arabic link, but notes imports can fail when files contain unrecognized characters, forcing manual entry link. Workday warns that parsing results vary by resume format and word order and that its parser does not auto-fill Languages and Skills link.

For bilingual Arabic/English CVs, this creates an audit task most teams skip. Check if the Arabic and English versions of an employer, city, or skill survive parsing, and also check for transliteration issues. If one version says الرياض and the other says Riyadh, the extracted fields may be uneven, even with a polished CV.

Stage 2: The hard filters that reject someone before ranking starts

A knockout question, rather than the AI itself, is often responsible for rejecting a candidate.

Automatic rejections often happen before any scoring is calculated, triggered by form rules. Workable documents auto-disqualify rules for work authorization eligibility, location, minimum qualifications/experience, and salary expectations link. Gem describes knockout questions for enforcing requirements such as work eligibility and required certifications or licenses link.

The screening stack generally follows this sequence:

1. Parse the CV.
2. Read form answers.
3. Apply knockout logic.
4. Rank the survivors.

If your job requires local presence in Dubai, a forklift certificate, or a specific salary cap, the candidate can be stopped cold by those answers alone. This stage is the cleanest place for a team to audit because the rules are explicit and the consequences are immediate. The failure is often self-inflicted, stemming from a location rule written too tightly or an experience threshold that screens out capable people.

Stage 3: How matching and ranking work in practice

Once a candidate clears the knockout filters, the system compares the extracted CV data to the job description.

That comparison is a blend of several factors, including:

• required criteria
• preferred criteria
• keyword matching
• semantic matching
• evidence from titles, dates, education, and skills

Workable’s Screening Assistant shows how well a candidate matches a job’s hard requirements and can display an AI rating link. Workday exposes a Candidate Skills Match score—Strong, Good, Fair, Low, Pending, and Unable to Score—on the requisition grid link. It also lets recruiters drill into score details for education, experience, and specific skills link.

Aligning the CV with the job description is a straightforward process: if the job asks for warehouse supervision, a GCC driving license, and Arabic customer handling, the system looks for that evidence in the extracted profile. Strong bullets with dates, results, and recognizable titles survive this handoff better than stylish but thin formatting.

Stage 4: What the recruiter sees on screen

Recruiters start with a summary card, not the full CV.

That card is built from whatever survived parsing, extraction, and scoring. Workable states recruiters get short summaries of candidate profiles, match information, and an AI rating link. In Workday, the candidate list can show the Candidate Skills Match score first link.

Greenhouse has a different shape. Its candidate profile centers name, pronouns, contact info, time zone, and tags. The main panel emphasizes stages, scorecards, scheduling, and tasks link. The deeper material—custom questions, documents, and application history—is in the application review section link.

This summary functions as the real shortlist desk; if the location field is blank, the skills field is thin, or the hard requirement checklist shows a single missing item, the candidate never earns a full read.

Stage 5: Side-by-side — Plain, Formatted, and Bilingual CVs

Run the same candidate through in three versions, and the differences appear fast.

CV version	Raw parsed text	Extracted fields	Ranking impact	Common loss point
Plain English CV, single column	Clean reading order	Strong extraction of titles, dates, skills, location	Better keyword matching and job description alignment	Low
Heavily formatted CV with tables/columns/header contact info	Text may be split or reordered; headers/footers may vanish	Contact, dates, or section content can fragment	Score drops because required evidence is missing or misplaced	High
Bilingual Arabic/English CV	Parsing may succeed, but mixed scripts need checking	Arabic may parse, but field consistency varies by format, word order, and character handling	English side may carry more of the ranking if the job description and knockout questions are English-only	Medium to high

The bilingual case is particularly important for several reasons:

• Arabic parsing support exists in major systems like Greenhouse link.
• Parsing quality still varies by file structure and word order link, link.
• Candidate-facing language support is not automatic translation. Greenhouse separates application language from job content translation, which must be entered manually link.

This final point is significant: a team can offer an Arabic application flow while screening against English-only hard requirements.

The one takeaway for SMB teams: audit your screening workflow

Instead of arguing about whether the ATS or the AI is “good,” teams should inspect their workflow.

Inspect one recent rejected application by examining each stage of the process:

• the original CV
• the raw parsed text
• the extracted fields
• the knockout-question answers
• the ranking/scoring output
• the recruiter-facing summary

If the failure starts at parsing, fix your CV guidance. If it starts at knockout questions, rewrite the rule. If it starts at ranking, tighten the job description. If it hits bilingual applicants hardest, review Arabic and English fields side by side.

The journey of a CV through the system is inspectable. Teams should open the last rejected file, follow the trail, and use the facts they find to improve their hiring process.