What "AI application review" actually does

The category goes by several names: AI resume screening, application triage, AI CV scoring, automated screening. The underlying capability is the same: read an application, score it against the role, produce a shortlist + rationale.

Modern AI application review has four sub-capabilities:

1. Parse the application. CV, cover letter, custom-question answers, portfolio links. Extract structured fields: employment history, education, certifications, skills, location, salary expectations.
2. Score against the role. Skills match, experience relevance, seniority fit, location match, salary band match. Output: a fit score with breakdown by dimension.
3. Detect quality issues. AI-generated text, fraud signals (impossible tenure, fabricated certs), inconsistencies (claimed skills not supported by experience).
4. Recommend an action. Advance, queue for review, reject with reason. Best-in-class tools require human confirmation before any reject — the EU AI Act Article 14 oversight requirement.

How AI scores a CV — the technical breakdown

Most vendor marketing presents AI screening as magic. It's not — it's a stack of well-understood techniques. Understanding the stack helps you evaluate vendors and audit outputs.

Layer 1: Document parsing

PDF / DOCX / image-based CVs go through a parser that extracts text + structure. Quality varies wildly. Modern parsers handle most layouts cleanly; older ones break on multi-column or graphic-heavy CVs. Test with weird CVs before trusting the vendor.

Layer 2: Entity extraction

An NER (named-entity-recognition) model identifies companies, roles, dates, skills, locations, certifications. This is where the structured data is built.

Layer 3: Semantic embedding

The full CV gets turned into a dense vector via an embedding model (OpenAI ada-002, Cohere, proprietary). The job description gets the same treatment. Cosine similarity between the two = semantic match score.

Layer 4: Explicit feature matching

Hard requirements (years of experience, must-have skills, location, work authorization) get checked deterministically. A candidate without legal work authorization in the EU shouldn't get a high score regardless of semantic match.

Layer 5: LLM reasoning

An LLM (Claude, GPT-4) reviews the structured data + semantic score + explicit features and produces the final score with reasoning. "Strong match on technical depth, weak match on years (4y vs 6y required), strong on language requirements." This is also where bias audits happen.

Layer 6: Bias audit + Article 14 confirmation

The system tracks adverse-impact metrics across protected characteristics (4/5 rule). It also requires explicit human confirmation before any AI-driven reject is committed to the candidate record. This is the Article 14 human-oversight requirement.

EU AI Act + GDPR — what "compliant screening" means

AI screening is the clearest case of high-risk AI in recruiting. It directly determines which candidates progress, which makes it textbook Annex III. Compliance obligations are extensive.

EU AI Act obligations (effective August 2026)

• Article 11 — Technical documentation: the vendor must publish technical documentation including training data, performance metrics, limitations, known failure modes.
• Article 12 — Automatic logging: every screening decision must be logged with timestamp, model version, input data hash, output score, and the human action taken.
• Article 13 — Transparency: candidates must be informed that AI is used in their evaluation (depending on jurisdiction interpretation; EU consensus is yes).
• Article 14 — Meaningful human oversight: the recruiter must be able to review, override, and reject the AI's recommendation. UX matters: a "click to confirm" rubber-stamp button does not meet the threshold; the recruiter must see the reasoning and have the ability to disagree.
• Article 26 — Documented oversight: deployers must maintain a written SOP.
• Article 27 — Fundamental Rights Impact Assessment: deployers in public sector or certain regulated industries must complete an FRIA.

GDPR obligations (independently)

• Article 6 — Lawful basis: processing candidate data for AI scoring requires lawful basis. Legitimate interest is typical; consent is best practice.
• Article 13/14 — Information to candidates: the candidate-facing privacy notice must disclose AI scoring, the logic involved, and the significance/consequences.
• Article 22 — Solely automated decision-making: if the AI's decision is final without human review, Article 22 applies — and Article 22 grants candidates the right to human intervention, to express their point of view, and to contest the decision.
• Article 35 — DPIA: required for high-risk processing.

What "compliant" actually looks like in UX

A compliant AI screening workflow has these properties:

• Candidate-facing notice in the application form disclosing AI scoring
• Each AI score visible to the recruiter with reasoning breakdown (not a black-box number)
• Explicit "confirm reject" UX with reasoning required — not auto-applied
• Bias-audit dashboard tracking adverse impact across protected characteristics
• Article 12 audit log queryable for any candidate decision
• Article 22-compliant appeal process: a candidate who's been rejected can request human review

tenperzent ships this UX today. Most US-led tools support the back-end logging but the UX rubber-stamps reviews — which won't survive a regulator audit.

Top tools compared

tenperzent — full breakdown

Semantic CV screening with transparent scoring (skill match %, experience relevance, semantic similarity, seniority match — each visible). Real-time bias audit. Article 14 confirm-reject UX. Article 12 audit log accessible to admin users. Multi-language support (50+ including IT/DE/FR/ES). Included in Business plan at €89/recruiter/month.

Metaview Triage — full breakdown

Strong on UX, similar capabilities to tenperzent's screening, less mature on the Art 14 confirm UX (more rubber-stamp friendly). Pricing per-resume reviewed ($1-3), which can scale fast on high-volume roles. No published Art 11-26 pack as of Q2 2026.

HireEZ Applicant Review — full breakdown

Built into the HireEZ sourcing platform. Strong on the candidate-from-sourcing-to-screening flow. Less mature on stand-alone applicant screening if you're not already a HireEZ customer.

How to deploy AI screening without breaking things

Phase 1 — Privacy + DPIA (Week 1)

• Update candidate-facing privacy notice — disclose AI screening, lawful basis, significance
• Run DPIA with your DPO
• Document oversight procedure (Article 26 SOP)
• Run FRIA if required for your sector

Phase 2 — Pilot on 2 roles (Week 2-3)

• Pick 2 high-volume roles (engineering, sales)
• Run AI screening in parallel with manual screening for 50 candidates per role
• Compare AI shortlist vs human shortlist — overlap should be 60-75% for a well-calibrated tool
• Investigate disagreements: was the AI right, or was the human right?

Phase 3 — Calibrate (Week 4)

• Adjust skill weightings based on Phase 2 findings
• Add custom must-have rules where the AI was missing nuance
• Tune the human-confirm UX so recruiters actually read the reasoning
• Set up bias-audit alerting (4/5 rule across gender, ethnicity-proxy, age-proxy)

Phase 4 — Roll out (Month 2)

• Expand to all high-volume roles
• Monitor: recruiter override rate (should be 15-30%; below 10% = rubber-stamping, above 40% = miscalibrated AI)
• Weekly review of edge cases

Phase 5 — Optimize (ongoing)

• Quarterly bias-audit review
• Quarterly DPIA refresh
• Monitor candidate-complaint rate; Article 22 appeals should be rare but supported

Pricing figures reflect public listings, market estimates, and customer reports as of Q2 2026. Vendors change pricing without notice — always verify directly with the vendor for current quotes. tenperzent publishes transparent pricing at tenperzent.com/pricing.