AI CV screening in the UAE: write the rules, then let good candidates through

AI strategy & adoption

A single Dubai job advert can pull several hundred CVs in days, so employers switch on AI screening inside the ATS or buy a GCC assessment tool, often before anyone has written down what a good candidate looks like. The result is a filter that quietly drops the people you needed. Here is how to define the scoring rules first, test them against your past hires, put a person on every rejection near the cut-off, and hold a vendor to the standards the UAE already expects.

A single Dubai job advert can pull several hundred CVs in days. Recruiters in the UAE told the Khaleej Times in 2026 that one vacancy now attracts "several hundred applications within a matter of days", and one founder who spoke to the same publication put numbers on the shift: a role that used to draw 200 applications now draws 2,000. The response has been predictable: switch on the AI screening already sitting inside the applicant tracking system (ATS), or buy an assessment tool built for the Gulf. What far fewer employers do is decide what "a good candidate" means for that specific role, write it down, and check who the software is filtering out. That gap is where good candidates disappear, and where the UAE's data protection law starts to apply to a hiring decision.

What AI CV screening is actually doing to your vacancy

Three layers do the filtering in a typical UAE hiring stack today, and they behave differently. A 2026 review of AI assessment tools for UAE hiring compares the main options now on the market.

1. Parsing and ranking in the ATS. The careers page or job board feed receives the CV, extracts name, work history, skills and education, matches keywords from the job description and ranks candidates. MakeMyCV's 2026 guide to ATS use in Dubai describes the failure modes plainly: multi-column layouts, graphics, images saved as JPEG, key details in headers and footers, and non-standard headings all break the parse. The important nuance is that an ATS usually does not auto-reject. It ranks, and a CV ranked low is never opened, which has exactly the same effect as a rejection with no human ever making a decision.
2. Assessment platforms. Tools such as Evalufy sit alongside the ATS and run structured tests (cognitive, behavioural, technical, language) plus AI-scored video responses, producing scorecards before a recruiter spends time.
3. Conversational screening. Newer GCC-built tools such as OVI's Milo agent run an audio conversation with the candidate against a rubric and score the transcript. Some tools score video, expressions or tone instead. That design choice has legal consequences in the UAE, covered below.

The pressure behind all three is volume against the number of CVs a person can actually read. When a role draws hundreds of applicants and a recruiter gets through perhaps 20 an hour, most CVs are never seen. That is where the cost of a wrong rejection rises: the candidate you filtered out takes an offer elsewhere and you never know you lost them.

There is a quieter problem too. Avinav Nigam of TERN Group told the Khaleej Times that many screening systems "prioritise keywords over actual capability", so a candidate who knows how to optimise a profile for an algorithm outranks someone better at the job. If your screening is still a keyword match against the job description, you are selecting for CV-writing skill. Semantic matching helps; it does not fix the underlying issue, which is that nobody has defined what good means, so the tool infers it from the advert.

The UAE rules that already apply to an automated rejection

The UAE has no separate "AI hiring law", but two existing instruments cover most of what matters, and a third is worth following as the standard for how to evaluate these tools.

Federal Decree-Law No. 45 of 2021 on Personal Data Protection (the PDPL). In employment, a rejection is a decision with real impact on a person, so treat the screening step as automated processing of personal data with legal consequence. Three articles matter:

• Article 13 gives the data subject the right to know the types of data processed, the purposes, and the decisions made based on automated processing, including profiling.
• Article 18 gives the right to object to decisions resulting from automated processing that have legal impact or adversely affect the person, and Article 18(4) obliges the controller to include a human element in reviewing those decisions at the data subject's request.
• Article 21 requires a data protection impact assessment before processing that involves a systematic and comprehensive assessment of people using automated processing or profiling where there are legal consequences or serious impact, and requires you to review that assessment regularly.

In practice that means three duties before you enable auto-scoring on a live vacancy: tell candidates the software is used and how it works, keep a working human review route, and run a DPIA you revisit whenever the rules change.

Federal Authority for Government Human Resources (FAHR) guidance. FAHR's guidelines on the use and evaluation of AI in talent assessment, produced with SHL, are written for federal government entities but are the clearest UAE-side statement of how to evaluate an AI assessment. The headline position is that no AI assessment should make a decision without human oversight. The framework turns on six questions (training data, transparency, bias, validity, finality, candidate disclosure) and flags three risk classes: legal risk from data protection and bias; public relations risk from candidates who feel the process was unfair or intrusive; and effectiveness risk from a tool that does not work on your population. Those six questions are the spine of the vendor checklist below.

Emiratisation. AI can rank CVs; it does not understand Emiratisation obligations, visa eligibility or MOHRE reporting. As Innovations Global's 2026 employer guide notes, firms with 50 or more employees face recurring penalties for unfilled Emirati quota positions, and that exposure does not disappear because sourcing was automated. Screening rules that penalise career breaks or non-linear CVs can quietly work against Emirati candidates. Check.

Write the scoring rules before you switch the tool on

The scoring rules belong to the hiring manager, not to the vendor and not to HR alone. Software can only implement what somebody wrote down. Here is the minimum artefact you need, in the order you need it.

1. Gates before weights. Split criteria into pass/fail gates (eligibility, licence requirements, minimum experience, availability) and weighted criteria (depth of relevant experience, domain, named systems, evidence of results). A weighted score should never compensate for failing a legal or eligibility gate.

2. Weights that add to 100, in a file anyone can read. Keep the rubric as configuration, versioned per role, so you can show a candidate or a regulator exactly what was applied. Example:

# screening-rules.yaml: Fleet Operations Supervisor, Dubai
gates:
  - eligibility: visa_status in [residence, golden, transferable] or uae_national
  - availability_within_days: 30
  - relevant_experience_years: ">= 2"
  - cv_parse_quality:
      rule: ">= 0.95"
      on_fail: manual_queue   # never auto-reject on a bad parse
weights:
  relevant_experience_years: 25
  freight_or_3pl_last_role: 20
  named_wms_tms_systems: 15
  measurable_results_in_cv: 15
  currently_based_in_uae: 15
  tenure_stability: 10
cut_bands:
  auto_advance_min: 70        # person still confirms
  human_reads_every_cv: 55    # 55-69: a person reads it
  auto_reject_below: 55
audit:
  random_reject_sample: 0.10
  recalibrate_if_override_rate_above: 0.10
  appeal_sla_working_days: 5
  appeal_mailbox: "careers-review@yourcompany.ae"

3. Reason codes, not just scores. Every CV should carry a score plus the reason codes that produced it (missing: named systems; gap: no logistics experience in the last role). This is what lets you answer an appeal, satisfy Article 13, and spot a systematic error such as the parser failing on Arabic-language CVs.

4. Candidate disclosure. Put it on the application form, in English and Arabic:

When you apply, our recruitment software scores your CV against criteria set by the hiring manager for this role. Our recruitment team reviews the results. You can ask a person to review your application at any time by writing to careers-review@yourcompany.ae.

5. A cut-off band, not a cut-off line. The single most important design choice. Any candidate within a defined margin of the cut gets a human read of the full CV. Rejections far below the cut can be automated with reason codes and sampling; rejections near the cut cannot. This is what "letting good candidates through" means in operational terms.

Worked example: 412 CVs for one Dubai logistics role

A Dubai logistics employer posts a Fleet Operations Supervisor role. 412 CVs arrive in 9 days. The numbers below are illustrative; the sequence is the part that matters.

Step 1: the hiring manager writes the rubric (half a day). The panel agrees gates (right to work, availability within 30 days, 2+ years in logistics customer service, CV must parse cleanly) and the weights in the config above. "Measurable results in CV" is weighted 15 deliberately, because it is harder to game than keywords.

Step 2: back-test on past hires before go-live (one day). Pull the last 12 people hired into this role and 48 CVs rejected in past rounds. Run the rules over all 60.

• Version 1 scored only 8 of the 12 past hires above the cut: 67 per cent recall. Unacceptable. Below 90 per cent, your rules describe something other than your own definition of good.
• Diagnosing the misses: three hires had image-based or multi-column CVs that parsed badly (they came in by referral, so nobody noticed) and one had short average tenure but strong performance. Fixes: a bad parse sends the CV to the manual queue instead of scoring it; tenure weight drops from 15 to 10; referrals and internal applicants are hard-coded to always get human review.
• Version 2 scores 11 of 12 past hires above the cut (92 per cent). The one miss is the internal transfer, which the new rule catches. Of the 48 previously rejected CVs, 5 now land in the human-review band, which tells you the old manual process was dropping people at the margin too.

Step 3: run the first live round with bands. Of 412 CVs: 55 score 70 or above (auto-advance, recruiter still confirms each one), 88 fall in the 55 to 69 band (a person reads every full CV), 269 score below 55 (auto-reject with reason codes, 10 per cent sample re-read).

Step 4: check the time maths. Reading all 412 CVs at 3 to 4 minutes each is about 24 hours of recruiter time. The new process reads 143 CVs plus 27 audit reads: about 10 hours. You save roughly 60 per cent of screening time while making the deliberate, budgeted choice that 88 borderline CVs get human judgement. That band is the cost of not losing good candidates. Do not let anyone "optimise" it away in round three.

Check every round: who got filtered out

The rules are a hypothesis about your talent market. Each round, test the hypothesis with a short audit pack.

Keep the audit pack for at least the recruitment cycle. If a candidate objects under Article 18 and asks for human review, your five-working-day SLA and your reason codes turn that request into a 20-minute task instead of a panic.

The questions to ask a vendor before it touches a live vacancy

FAHR's six questions are the right spine for any demo or RFP, in any sector. Add the data and commercial questions underneath.

1. Training data (FAHR Q1). Was the model trained on CVs from this region, in which languages and dialects, for which occupations? Ask what happens to an Arabic-language CV, or to an Egyptian, Levantine or Khaleeji accent if audio is scored. Ask whether the training data included outcomes (who got hired, who performed) or only CV text.
2. Transparency (Q2). Can you show, per candidate, which inputs drove the score? Ask for the technical manual. If the answer is a black box with a single "fit score", you cannot explain a rejection to a candidate or to a regulator.
3. Bias (Q3). What adverse impact analysis has been run, by which groups, and what were the results? What does the tool do when demographic data is missing (it usually is)?
4. Validity (Q4). What validation evidence exists, and is it content, construct or criterion related? Ask the vendor to validate against your historical hires before you trust the scores.
5. Finality (Q5). Can we configure cut scores? Can recruiters override? Is auto-rejection on or off by default? FAHR's position is that AI assessments should inform human decisions, not make them.
6. Candidate disclosure (Q6). Do you supply English and Arabic notice text, and where must it appear?
7. Data. Where is candidate data hosted, for how long, who are the sub-processors, is our data used to train your models, and can we export and delete everything at exit? Ask how the tool supports the Article 21 impact assessment. Claims such as "PDPL compliant" need a document behind them, not a website badge.
8. Commercial. What is the cost at our volume, what counts as a billable unit, and what does exit look like?

What the market looks like in 2026

Prices and posture as published or reported this year. Several figures below come from vendors writing about their competitors, so verify everything in writing before signing.

A few observations from reading the material closely. Evalufy publishes a genuine free tier (15 assessments per month, no credit card), enough to test on a live requisition, and claims PDPL compliance with ISO 27001-certified MENA data centres: ask for the certificates. Published descriptions of Evalufy's headquarters also disagree between reviews, which is exactly the kind of detail to confirm with the vendor rather than assume. OVI's comparison of the four tools is useful but it is OVI's own material, and OVI sells a competing product, so treat its pricing for HireVue and SHL as a starting point for quotes rather than fact. SHL co-authored the FAHR guidance, which is a fair reason to include it on a shortlist for psychometric-heavy roles. And notice the design split that matters most under the PDPL: transcript-based scoring (what the candidate said) carries far less risk than tools scoring face, voice or expression, which is closer to sensitive personal data and a heavier Article 21 exercise.

What to do next

1. Pick one live or upcoming vacancy and write the gates and weights. Half a day, hiring manager owns it, HR facilitates.
2. Back-test the rules on your last 12 hires and a sample of rejected CVs. Do not go live below 90 per cent recall.
3. Configure cut bands, with a person reading every CV within 15 points of the cut, plus a 10 per cent audit sample of auto-rejections.
4. Turn auto-rejection off by default in your ATS until the bands and reason codes exist. Most systems will let you rank without rejecting.
5. Run the data protection impact assessment (PDPL Article 21) and put the disclosure text and appeal mailbox on the application form, in English and Arabic.
6. Put the six FAHR questions plus the data questions in writing to every vendor, and require answers before they see a single live CV.
7. Review the audit pack after round one and recalibrate anything the triggers flag.
8. Keep the rules per role, versioned, and revisit them whenever the role or the market changes.

If your leadership team cannot agree the rubric for a role, that is a strategy problem before it is a tooling problem. Our AI strategy work, run by our own team in the UAE, starts exactly there: we look at how your business runs, from processes and systems to data, tell you plainly where AI will pay off, what it takes and what to do first. Screening is one of the highest-volume, highest-risk AI decisions most UAE employers run, and it is worth getting right before you scale it.

https://www.azrty.com/blog/ai-cv-screening-in-the-uae-write-the-rules-then-let-good-candidates-through

<a href="https://www.azrty.com/blog/ai-cv-screening-in-the-uae-write-the-rules-then-let-good-candidates-through">AI CV screening in the UAE: write the rules, then let good candidates through</a> (Azrty)