Built to Minimise Bias.
Designed for Enterprise Accountability.
Skima AI applies structured fairness evaluation and privacy-first AI architecture to help large enterprises screen candidates equitably at scale, with the transparency your legal and HR teams require.
The Stakes
The Legal and Reputational Risk is Real
AI hiring tools are under scrutiny in the EU, the US, and various cities and states. Using an unvalidated screening system exposes your organization and, in some areas, your software vendor, to legal risks.
EU AI Act · Aug 2026
AI hiring tools are classified as high-risk systems under Annex III. Requires documented bias evaluation, human oversight, and ongoing monitoring.
Up to €35M or 7% revenue
EEOC · Title VII · US
Disparate impact doctrine applies to algorithmic screening. Mobley v. Workday (2025) certified that AI vendors can be held liable as employer agents.
Class action exposure
NYC LL144 · California ADS
NYC requires annual third-party bias audits and candidate disclosure for automated employment decision tools. California's ADS regulations took effect Oct 2025.
Mandatory annual audits
This page documents Skima AI's architecture, evaluation methodology, and data governance, which are designed to help your organisation use AI-assisted screening with confidence, along with the documentation trail that regulators and auditors expect.
Architecture First
Fairness Built Into the Model, Not Bolted On
Most AI screening tools apply post-hoc bias filters. Skima AI's approach is different: our architecture structurally eliminates the most common sources of systematic bias before a single resume is scored.
Job-Related Scoring Signals
Skima AI scores candidates on verified skills, role-relevant experience, qualification match, and tenure patterns. No proxy variables (university prestige, name-based inference, address) are used.
Signal HygieneProtected Attribute Exclusion
By contractual DPA obligation, GDPR Article 9 special category data (including health, religion, racial origin, and disability) is never processed as a scoring input under any circumstance.
GDPR Art. 9 CompliantInternal-Only PII Processing
Candidate resumes and PII are processed exclusively via Skima AI's proprietary internal models hosted in our Dublin, Ireland data centre. No candidate data is ever sent to public LLM APIs.
Privacy-FirstHuman-in-the-Loop Always
Skima AI's AI ranks and surfaces candidates; it never makes final hiring decisions. Every progression, rejection, and offer requires human authorization. AI is an input, not the decision-maker.
EU AI Act CompliantTesting Methodology
How We Evaluate for Bias
Skima AI conducts structured bias evaluations using the Disparate Impact (Four-Fifths Rule) standard as specified in the EEOC Uniform Guidelines on Employee Selection Procedures (29 CFR Part 1607).
Disparate Impact Analysis
We calculate the selection rate for each demographic group and compute impact ratios relative to the highest-selected group (the reference group). Any group whose selection rate falls below 80% of the reference group's rate (an impact ratio below 0.80) is flagged for investigation and model review, per the Four-Fifths Rule.
Synthetic Profile Testing
We construct matched candidate pairs identical in qualifications, skills, and experience, varied only in demographic signals (names, universities, demographic disclosure language). These paired profiles test whether the model's scoring is driven by job-relevant signals or demographic proxies.
Intersectional Analysis
We test compound demographic combinations (e.g., Black or African American Female, Asian Male) to detect adverse impact that may not appear in single-category analysis but emerges at the intersection of multiple protected characteristics.
Protected Classes Evaluated
Our evaluation covers all EEOC-protected classes relevant to hiring: Sex, Race and Ethnicity (Asian, Black or African American, Hispanic or Latino, White), Age (Above 40 / Below 40, aligned with ADEA), Disability Status, and intersectional combinations thereof.
Meets the Four-Fifths Rule. No significant adverse impact detected. Model operates within the acceptable range per EEOC Uniform Guidelines.
Potential adverse impact detected. Model is flagged for review, root-cause analysis, and possible retraining before continued deployment.
Significant adverse impact. Model deployment is paused for that use case until remediation is confirmed and re-evaluated.
Fairness Evaluation
Internal Bias Evaluation Results
Results from Skima AI's internal fairness evaluation of the core candidate-screening model. Evaluated across 792 profiles spanning 5 demographic splits, using the Disparate Impact (Four-Fifths Rule) methodology per the EEOC Uniform Guidelines on Employee Selection Procedures.
Sex
ClearImpact ratios calculated relative to highest-selected group (Female, 54.55%). All groups above 0.80 threshold.
Race / Ethnicity
ClearImpact ratios calculated relative to highest-selected group (Hispanic or Latino, 58.08%). All groups above 0.80 threshold.
Age (Above 40 / Below 40)
ClearImpact ratios calculated relative to highest-selected group (Below 40, 61.31%). All groups above 0.80 threshold.
Disability Status
ClearImpact ratios calculated relative to highest-selected group (No Disability, 54.23%). All groups above 0.80 threshold.
Overall Evaluation Summary
All ClearAll 5 demographic splits return impact ratios above the 0.80 Four-Fifths Rule threshold across 792 evaluated profiles. Zero groups breach the concern threshold. Results are consistent with Skima AI's architectural commitment to evaluating candidates exclusively on job-relevant qualification signals. Full methodology documentation and raw evaluation data are available to enterprise clients upon request under NDA.
Intersectional Bias Analysis
Race/Ethnicity × Sex compound combinations · Reference group: Hispanic or Latino / Female (59.60% selection rate, impact ratio 1.0000)
| Race / Ethnicity | Sex | Sample Size | Selection Rate | Impact Ratio | Status |
|---|---|---|---|---|---|
| Hispanic or Latino | Female | 99 | 59.60% | 1.0000 | ClearReference |
| Hispanic or Latino | Male | 99 | 56.57% | 0.9492 | Clear |
| Asian | Female | 99 | 54.55% | 0.9153 | Clear |
| Asian | Male | 99 | 53.54% | 0.8983 | Clear |
| Black or African American | Male | 99 | 53.54% | 0.8983 | Clear |
| White | Female | 99 | 52.53% | 0.8814 | Clear |
| White | Male | 99 | 51.52% | 0.8644 | Clear |
| Black or African American | Female | 99 | 51.52% | 0.8644 | Clear |
Regulatory Alignment
Compliance Coverage by Region
Skima AI is designed to support your compliance obligations across key regulatory frameworks. The deploying employer bears ultimate regulatory responsibility; Skima AI's role is to provide the architecture, documentation, and controls that make that compliance achievable.
- AI hiring tools classified as high-risk: Skima AI maintains documented risk management processes per Articles 8–15
- Bias evaluation data available to support Article 10 data governance obligations
- Human oversight at every decision point: no autonomous hiring decisions
- Technical documentation available for regulatory inspection
- Enforcement date: August 2, 2026. Skima AI's controls are designed to be ready ahead of this
- Full Controller-to-Processor DPA with EU Standard Contractual Clauses (Module 2)
- GDPR Article 9 special category data (health, religion, racial origin) contractually prohibited as scoring inputs
- Data subject rights (access, erasure, portability) supported via documented DPO process
- PII processed exclusively via internal models in Dublin, Ireland, never sent to public LLMs
- Privacy by Design and Privacy Impact Assessment processes in place
- Disparate Impact (Four-Fifths Rule) evaluation per EEOC Uniform Guidelines covers all protected classes: race, sex, national origin, religion, age, disability
- Scoring signals are strictly job-relevant: no proxy variables that correlate with protected characteristics
- Human reviewer authorises all candidate progression: no automated rejection without human oversight
- Documentation trail maintained to support EEOC investigation response
- Model does not use resume signals that constitute illegal pre-employment inquiries
- NYC LL144: Skima AI provides bias evaluation data, methodology documentation, and audit-trail exports to support the annual third-party audit requirement
- NYC LL144: Architecture supports candidate disclosure notification workflows
- California ADS (FEHA, eff. Oct 2025): Skima AI's evaluation documentation supports automated decision system impact assessments
- Illinois AI Video Interview Act: AI Video Agent collects candidate consent, supports data deletion requests, and avoids facial analysis scoring
- Note: LL144 compliance responsibility lies with the deploying employer. Skima AI provides the tooling to make it achievable.
Data Architecture
Privacy-First, By Design
Skima AI's data processing architecture is structured so that candidate privacy and model isolation are enforced at the infrastructure level, not just by policy.
- Candidate resumes and PII are processed exclusively by Skima AI's proprietary internal models, hosted in AWS Dublin (EU)
- Public LLM APIs (e.g. OpenAI) are used only for non-PII tasks like job description generation; candidate data never crosses this boundary
- Each enterprise client operates in a logically isolated tenant environment. No candidate data is ever aggregated across clients
- DPA contractually prohibits using candidate data to train or improve Skima AI's generalised base models
- All data in transit encrypted via TLS. Data at rest encrypted at the storage layer. Multi-factor authentication enforced on all access
Candidate PII → Public LLM APIs. Prohibited by Skima AI DPA · Never executed
Human Oversight
AI Informs. Humans Decide.
Skima AI operates as an intelligence layer that surfaces, ranks, and contextualizes candidates. At no stage does the AI make or execute a hiring decision. Every consequential action requires human authorization.
Resume Screening
AI scores and ranks candidates against the job requirements. Shortlist suggested, not enforced.
AI AssistsShortlist Review
Recruiter reviews AI rankings, applies judgement, and approves which candidates advance. AI score is one input among several.
Human DecidesInterview Stage
AI Video Agent conducts structured L1 screening with candidate consent. Transcript and signals surfaced for human review.
AI AssistsHire / Reject
Final hiring decision is made and recorded exclusively by a human hiring manager. No automated rejection without human sign-off.
Human DecidesContinuous Improvement
Bias Evaluation is Ongoing: Not a One-Time Test
Fairness is not a certification achieved once and forgotten. Skima AI maintains a structured monitoring programme to detect drift, respond to new regulatory guidance, and evolve with best practice.
Internal Fairness Review
Bias evaluation runs are conducted quarterly using updated synthetic profiles. Results are reviewed by Skima AI's model team. Any impact ratio approaching the Review band triggers a root-cause investigation.
Security & Compliance Audit
An independent third-party conducts an annual risk assessment across all systems processing customer personal data. Results inform Skima AI's risk treatment programme and are available to enterprise clients.
Regulatory Tracking
Skima AI's legal and product teams actively monitor developments in AI hiring law (EU AI Act, EEOC guidance, new state-level regulations) to ensure architecture and documentation remain current.
Model Monitoring
Each of our models is periodically monitored for data distribution shifts. Significant data distribution shifts automatically trigger a review flag.
Independent Third-Party Audit
Skima AI is planning an independent third-party bias audit with a qualified HR compliance firm, aligned with NYC LL144 methodology. Enterprise clients will receive advance access to published results.
Client-Specific Documentation
Enterprise clients may request full evaluation methodology documentation, raw impact ratio data, and architecture diagrams under NDA for their own internal audit, legal, or procurement processes.
FAQs
What Enterprises Ask Us
Answers to the questions your legal, compliance, and HR leadership teams will ask before deployment.
Get Started
Screen at Scale.
Request our compliance documentation pack, or speak with our team about how Skima AI supports your specific regulatory environment.