Quick summary

Veranika Khlud's 2026 conceptual article develops a multi-layer governance architecture for algorithmic hiring that aims to translate the EU AI Act's high-risk obligations into organizational roles, processes, and controls. The paper argues fairness in hiring cannot be reduced to one-off audits or technical bias mitigation; instead, organizations need a decision architecture spanning strategy, data and models, human oversight, candidate-facing rights, and continuous learning.

Read the paper: https://doi.org/10.66972/ada21202624

---

What question does the study ask?

How can organizations govern "high-risk" AI systems used in recruitment so that compliance with the EU AI Act results in fair outcomes — not only technically, but procedurally and interpersonally — across the full recruitment lifecycle?

Method (how the author approached it)

This is a conceptual, theory-building paper. Khlud conducts a conceptual synthesis that brings together:

• the EU AI Act's requirements for high-risk systems;
• organizational justice theory (distributive, procedural, interpersonal, informational justice);
• algorithmic fairness and accountability research; and
• human resource management (HRM) governance literature.

From this abductive integration the paper derives a governance model (a six-layer decision architecture) and seven theoretical propositions intended to guide empirical work and practical implementation.

Main findings

The paper proposes a six-layer governance architecture for high-risk recruitment AI:

1. Strategic accountability — clear senior-level responsibility, allocation of decision rights, vendor governance and resourcing for AI-enabled hiring.
2. Lifecycle risk management — procurement controls, risk assessment and mitigation throughout design, deployment and decommissioning.
3. Data and model assurance — dataset documentation, provenance, bias testing, validation, and model documentation (e.g., model cards / datasheets).
4. Human oversight — meaningful human-in-the-loop roles with competence and authority to review and override automated outputs.
5. Candidate-facing transparency and redress — accessible explanations, notice that AI is used, clear appeal and correction channels for applicants.
6. Continuous monitoring and organizational learning — logging, post-deployment performance monitoring, periodic audits, and feedback loops into HR policies.

Beyond the architecture, Khlud formulates seven propositions explaining how specific governance mechanisms map onto aspects of organizational justice (distributive, procedural, interpersonal, informational).

Key conceptual conclusions:

• Fair recruitment requires integrating legal compliance, technical controls, and perceived fairness by candidates — not treating those as separate problems.
• Isolated audits, single bias tests, or disclosure statements alone are insufficient.
• Effective governance must allocate rights and responsibilities across vendors, deployers, HR teams, legal/compliance, and candidates.

How the study matters (practical implications)

For employers and vendors:

• Treat hiring AI as a socio-technical system: combine technical fairness checks with HR processes that preserve meaningful human judgment and candidate voice.
• Establish senior accountability and cross-functional teams (HR, legal, data science, compliance) responsible for lifecycle risk management.
• Implement robust data governance (datasheets/dataset provenance), model documentation, logging, and reproducible validation.
• Ensure human reviewers have the authority, access to information, and training needed to meaningfully intervene.
• Provide candidate-facing notices, plain-language explanations of how AI influences decisions, and effective redress channels.
• Monitor deployed systems continuously and feed operational findings back into procurement and HR policy.

For regulators and policymakers:

• The paper shows how EU AI Act obligations can be operationalized inside organizations; regulators should support guidance and checklists that reflect lifecycle governance rather than only technical metrics.

Limitations and caveats

• Conceptual only: the architecture and propositions are not empirically tested in field settings. Their real-world effectiveness, cost, and feasibility remain to be validated.
• Implementation complexity: the model assumes organizational capacity (cross-functional skills, resources) that may be limited in smaller employers or vendors.
• Evolving regulation and technology: the AI Act and vendor tools continue to change; governance frameworks must remain adaptable.
• Trade-offs and ambiguous metrics: the paper acknowledges trade-offs between fairness definitions (e.g., group parity vs. individual accuracy) but does not resolve which metric is appropriate in specific hiring contexts.
• Human oversight is not a panacea: oversight can be superficial if reviewers lack authority or are disconnected from model design; meaningful oversight requires training, access to logs, and clear decision rights.

What to do next (practical checklist)

For HR / Legal / AI teams evaluating hiring AI today, Khlud's architecture suggests this prioritized checklist:

• Assign senior accountability for AI hiring and set up a cross-functional governance body.
• Map your recruitment pipeline to identify where AI materially affects candidate outcomes.
• Require vendor documentation (model cards, dataset datasheets), logs, and evidence of bias testing as part of procurement.
• Define and test meaningful human oversight workflows and ensure reviewers can override automated outputs.
• Publish clear candidate notices and an accessible redress process; track candidate complaints and resolutions.
• Put in place continuous monitoring: operational metrics, fairness metrics over time, and periodic third-party audits where appropriate.

Next research steps

The paper's seven propositions are ripe for empirical testing. Useful next studies would:

• Field-test the six-layer architecture in organizations of different sizes and sectors to assess feasibility and impact.
• Evaluate whether candidate-facing redress mechanisms measurably improve perceived procedural justice and reduce appeals.
• Measure how different human oversight designs affect fairness outcomes and speed of hiring decisions.

---

Bottom line

Khlud's paper provides a practical, theory-grounded roadmap for turning the EU AI Act's requirements into organizational routines that meaningfully support fair algorithmic hiring. It shifts the conversation from single-point technical fixes to a lifecycle governance approach that ties legal compliance to organizational justice. The ideas are timely and useful to practitioners, but they need empirical testing and adaptation to fit varied organizational capacities.

Read the full article: https://doi.org/10.66972/ada21202624

---

Sources referenced in this draft

• Veranika Khlud (2026), "Governing Fair Algorithmic Hiring Under the EU AI Act: A Multi-Layer Decision Architecture for High-Risk Recruitment Systems," Applied Decision Analytics. https://doi.org/10.66972/ada21202624

()