The 82% figure and what it actually means

AI resume screening is now the norm, not the exception. 82% of hiring teams use some form of it. But "AI resume screening" covers everything from basic keyword matching to genuine machine learning models — and they perform very differently. They also carry very different bias risks, compliance obligations, and accuracy benchmarks.

Before evaluating whether your current AI screening setup is working — or before deciding whether to adopt AI screening — it's worth understanding exactly what category of technology you're actually using. The vendor label "AI-powered" is not diagnostic. The underlying mechanism is.

This article explains the actual technology at each level of the spectrum, where each approach works, and where each fails — including the specific candidate profiles that AI screening consistently mishandles.

The technology spectrum: four distinct approaches

Level 1: Keyword matching — the baseline, not AI

The most basic form of "AI" resume processing is keyword matching: the system scans resume text for the presence or absence of specific words or phrases. If the job description requires "Python" and "5+ years" and "AWS," the system searches each resume for those exact strings and scores based on the count of matches.

This is pattern matching, not machine learning. It's been in ATS software since the early 2000s. It should not be marketed as AI, but it often is. The failure modes are predictable and significant:

  • Synonyms fail. A candidate who writes "cloud infrastructure" when you searched for "AWS" scores zero despite being qualified.
  • Context is ignored. "Python — 3 years" and "managed a Python team for 3 years" score the same despite meaning very different things.
  • Formatting dependent. Keywords in tables, graphics, or certain PDF structures may not be read at all.
  • Gameable. Any candidate who knows to mirror the job description language will score well regardless of actual qualification.

If your ATS vendor cannot explain how their AI handles synonyms and contextual language, you likely have keyword matching, not ML — regardless of what it says in the marketing.

Level 2: ML-based resume parsing — the genuine table stakes

Modern ML-based resume parsing is qualitatively different from keyword matching. The technology uses trained models to understand the structure and meaning of resume text, extracting data into a standardised format regardless of how the resume is formatted.

What this means in practice: the parser reads a CV in any standard format and produces structured outputs — name, contact information, each job with company name, title, start and end dates, responsibilities described, each education entry with institution, degree, and dates, skills identified and normalised to a standard taxonomy. It does this reliably across different CV structures, layouts, and writing styles.

The 95%+ accuracy benchmark for modern parsers means: for 19 out of 20 standard CVs, the structured data extracted is complete and correct. The 5% failure rate is not random — it clusters in specific cases described in detail below. This accuracy level is high enough that ML-based parsing is genuinely useful as a data entry automation tool, freeing recruiters from manual data input for the vast majority of applications.

How to distinguish genuine ML parsing from keyword matching: ask the vendor to run the same CV through the system formatted two different ways — once with standard bullet points, once as a narrative paragraph. A genuine ML parser handles both correctly. A keyword matcher degrades on anything outside its expected format.

Level 3: Candidate scoring and ranking — where bias risk enters

AI candidate scoring goes beyond parsing to evaluation: the system assigns a numerical score or rank order to candidates based on how well their resume matches the target profile. This is where the technical and ethical complexity significantly increases.

Scoring models work by comparing candidate attributes (extracted from the parsed CV) against a target profile and producing a match score. The target profile can be defined in two ways:

Rules-based scoring: The recruiter or system administrator defines explicit rules — "candidate must have degree in computer science, 3+ years of Python experience, experience in a company of 100+ employees." The AI score reflects how many rules are met. This is more transparent and auditable than ML scoring, but rigid and limited.

ML-based scoring: The model is trained on historical data — typically past candidates and hiring outcomes — and learns to predict which candidates are likely to succeed based on patterns in that data. This is where bias risk becomes significant.

The bias mechanism is straightforward: if your past hiring data contains demographic patterns (intentional or structural), the ML model learns those patterns and reproduces them. A model trained on a tech company's past engineering hires — which skewed male and came from a narrow set of universities — will score future candidates accordingly, not because the model is explicitly discriminatory but because it has learned to replicate the patterns in its training data.

The practical implication is that ML scoring requires periodic bias auditing: demographic analysis of scoring outputs to identify whether any protected group is being systematically scored lower than the pattern of their actual qualifications would predict.

Level 4: Skills gap analysis — useful for specific contexts

Skills gap analysis compares the skills extracted from a candidate's CV against the skills required in the job description and quantifies the gap. Rather than a single score, it produces a structured breakdown: "candidate has 8 of 12 required skills, missing: Kubernetes, Terraform, security architecture."

This is genuinely useful for technical roles where skill coverage is measurable and important. For engineering roles with specific toolchain requirements, skills gap analysis provides a more interpretable and auditable output than a single score. The hiring manager can look at the gap and make a judgment about whether it's bridgeable.

The limitation: it only measures what can be stated in a CV. Soft skills, communication quality, leadership capability, and cultural fit cannot be assessed from resume text regardless of how sophisticated the AI is.

Where AI resume screening consistently fails

Unusual CV formats

Design CVs — heavily graphic, visually formatted resumes that designers, creative directors, and UX professionals often use — consistently underperform in AI parsing. The visual layout that makes them compelling to a human reader makes them difficult for text-extraction systems. Text embedded in graphics, custom fonts that don't parse cleanly, and multi-column layouts that disrupt reading order all reduce parsing accuracy.

The irony is specific: the candidates most likely to have visually impressive CVs — designers, creative professionals — are the candidates whose CVs are most likely to be misprocessed by AI screening systems.

Career changers

AI screening models score candidates based on matching their past experience to the target role profile. Career changers, by definition, have past experience in different roles — experience that contains transferable skills that don't map cleanly to the new role's keywords.

A marketing manager transitioning to product management has deeply relevant experience (customer understanding, cross-functional collaboration, communication, analytics) that an AI model trained to recognise product management vocabulary will not score highly. The transferable skills are real; the vocabulary match is poor.

This is one of the clearest cases where AI screening produces systematically wrong outcomes. Career changers with genuine transferable capability are routinely deprioritised in favour of candidates with directly matching (but potentially less capable) backgrounds. Human judgment is specifically valuable here.

Senior and executive candidates

Experienced executives often write brief, high-signal CVs that trust the reader to understand the significance of their roles and accomplishments. A 20-year career might be summarised in a page that looks "thin" to an AI model expecting detailed descriptions, tenure lengths in months, and explicit skill listings.

The specific failure mode: senior candidates who have moved past the need to demonstrate basic competencies in their CV text get scored lower than junior candidates who explicitly list those same competencies. The AI reads words; it cannot read the signal embedded in the brevity.

Roles requiring soft skill assessment

For any role where the key differentiating quality is something that cannot be stated in a CV — communication style, judgment under pressure, empathy, leadership presence, client relationship quality — AI resume screening provides limited value beyond basic eligibility checking. These qualities don't exist in resume text. Treating an AI screening score as a quality signal for these roles produces systematically wrong prioritisation.

AEDT compliance implications

The regulatory landscape for AI resume screening has shifted significantly since 2023, and the direction of travel is clear: more regulation, not less.

New York City Local Law 144 applies directly to AI tools used in employment decisions, including resume screening tools that "substantially assist or replace" human judgment. Requirements: annual independent bias audit, public disclosure of audit results, candidate notification. The law has been in effect since July 2023 with enforcement active.

EU AI Act: AI systems used in employment and worker management decisions are classified as high-risk under Annex III. This includes tools used for screening, ranking, and selection of candidates. High-risk classification requires: conformity assessment, technical documentation, ongoing monitoring, human oversight mechanisms. Enforcement is phased, with requirements for high-risk AI applications in full force from August 2026.

Illinois Artificial Intelligence Video Interview Act: Requires disclosure when AI is used to analyse video interviews and restricts sharing of video interview data. A signal of state-level regulation expanding beyond New York.

The practical implication for any company using AI resume scoring: document the AI's role in your process, ensure human review is genuinely in place for any rejection decision, conduct demographic analysis of scoring outputs at least annually, and be prepared to explain how an AI-influenced decision was made if a rejected candidate challenges it.

See exactly what Treegarden costs

All features included. Public pricing. No demo required to see the numbers. Startup: $299/mo · Growth: $499/mo · Scale: $899/mo.

View full pricing →

Practical guidance for using AI screening well

The honest framework for deploying AI resume screening responsibly in 2026:

Use ML parsing universally. The ROI on eliminating manual data entry is clear and the bias risk is minimal. Every candidate who applies should have their CV parsed automatically into a structured record.

Use AI scoring selectively. Restrict scoring to high-volume roles (100+ applications) with clearly defined criteria where the target profile is stable and consistent. Do not use scoring for senior roles, executive roles, roles where the criteria are evolving, or roles where the past hiring data reflects demographic patterns you're trying to change.

Audit scoring outputs before using them for decisions. Before deploying any scoring model to filter candidates, run it on a sample of 50–100 CVs and analyse the outputs by demographic proxy (where visible from name, university, previous employer). Identify whether any pattern suggests systematic bias against a protected group.

Maintain genuine human review. "Human review" that rubber-stamps AI scores is not human review — it's AI decision-making with human cover. The human review step needs to involve a recruiter actually reading CVs that scored at the borderline, not just approving the AI's top quartile.

Document the AI's role. For any role where AI screening influenced which candidates were reviewed, document this in your hiring records. The regulatory obligation to explain AI-influenced decisions is coming regardless of jurisdiction — building the habit now reduces risk.

Frequently asked questions

Is AI resume screening biased?

AI resume parsing (data extraction) carries minimal bias risk. AI candidate scoring and ranking carries significant bias risk because these models are trained on historical hiring data that may reflect past demographic patterns. If past hiring skewed in any demographic direction, the AI will reproduce those patterns. Bias auditing, demographic analysis of scoring outputs, and human review of rejection decisions are essential safeguards.

How accurate is AI resume parsing?

Modern ML-based parsers achieve 95%+ accuracy on standard CV formats in major languages. The 5% error rate clusters in specific cases: heavily graphic design CVs, academic CVs with unusual structures, CVs in languages with limited training data, and image-based scanned CVs. High enough accuracy that human spot-checking can focus on these edge cases rather than be applied universally.

Can candidates trick AI resume screening?

Yes, to a limited degree. Keyword stuffing and mirroring job description language can inflate scores in basic keyword-matching systems. Modern ML-based scoring models are harder to game because they look at semantic meaning, not just keyword presence. The more meaningful concern is that optimising for AI screening selects for candidates good at CV writing — not necessarily the best at the job.

Should I use AI to screen candidates?

AI resume parsing: yes always — reliable time saver, minimal downside. AI candidate scoring: use carefully, restricted to high-volume roles with clear criteria, with demographic analysis of outputs and human review of any rejection decision. Never as the sole mechanism for complex professional roles where individual candidate quality matters most.