AI for recruitment: how to speed up hiring

AI can reduce time-to-hire by 26%, but it is classified as high risk under the AI Act. We explain how to implement it with human oversight, fairness metrics, and a governance model that protects your company and candidates.

Artificial intelligence is transforming recruitment. Sixty-seven percent of organizations already use AI in their hiring processes, and that figure is expected to reach 80% before 2027. The numbers are tempting: a 26–50% reduction in time-to-hire, 30–50% savings in cost per hire, and an average ROI of 340% over 18 months.

But there is an uncomfortable reality that many companies ignore: AI used in recruitment is classified as high risk under the AI Act (Annex III). And for good reason. An algorithm that filters CVs, scores candidates, or prioritizes profiles is making decisions that directly affect the fundamental right of access to employment.

The best-known case is Amazon: its CV screening system systematically penalized women because the model had learned from historical data in which most hires were men. It was not an isolated failure: it was structural bias encoded in the algorithm.

In this article, we explain how to take advantage of AI in recruitment without falling into the bias trap, what the AI Act requires, which fairness metrics you should monitor, and how to implement a human oversight model that really works.

The promise: why AI in recruitment is so attractive

Traditional recruitment processes are slow, expensive, and, paradoxically, biased too. A human recruiter takes an average of 42–44 days to fill a position, and decisions are subject to cognitive biases (affinity, halo, and anchoring) that are difficult to detect and correct.

AI promises to address both problems:

Speed: An AI system can screen hundreds of CVs in minutes, identify the profiles that best match the role requirements, and reduce time-to-hire from 44 days to 11–26 days.
Cost: Recruiters spend up to 70% of their time on administrative tasks (screening, scheduling, and follow-up). Automating these tasks can reduce cost per hire by 30–50%.
Consistency: Unlike a human who evaluates differently at 9 a.m. than at 5 p.m., AI applies the same criteria to every candidate.
Reach: Tools such as Textio optimize job descriptions to remove biased language and attract a more diverse pool of candidates.

But AI consistency is a double-edged sword: if the criteria are biased, AI applies that bias with a precision and scale no human could match.

The risk: algorithmic bias in hiring

Algorithmic bias in recruitment is not a theoretical risk. There are multiple documented cases:

Amazon (2014): Its CV screening system penalized résumés that included the word “women” (for example, “captain of women’s chess club”) and favored typically masculine language. The system was withdrawn.
HireVue (2019–2025): It faced allegations of bias in its video interview analysis system. In 2025, the ACLU filed a complaint alleging discrimination against deaf and non-white candidates. HireVue removed facial analysis and moved to competency-based assessments.
Workday (Mobley v. Workday): The first major class action alleging racial, age, and disability discrimination in automated candidate screening.
iTutorGroup (2022): An EEOC lawsuit alleging age discrimination in its AI-powered hiring tool.

The most common types of bias in recruitment AI:

Gender bias: Models trained on historical data in which certain roles were dominated by one gender.
Racial bias: Algorithms that penalize names, universities, or ZIP codes correlated with ethnic groups.
Age bias: Systems that deprioritize profiles with more years of experience or older graduation dates.
Disability bias: Video or voice analysis that penalizes atypical patterns of speech, facial expression, or movement.

Key point: Bias is not always intentional or obvious. A model may not use the variable “gender” directly and still discriminate through proxy variables (type of university, hobbies, or CV writing style). Detecting it requires specific metrics.

What the AI Act requires: high-risk classification

The AI Act classifies AI systems used for recruitment as high risk (Annex III, employment category). This includes targeted job advertising, application screening, candidate evaluation, hiring or dismissal decisions, assigning tasks based on profiles, and performance monitoring.

The obligations that apply to you as the deployer:

Technical documentation: Document the purpose, data used, architecture, metrics, thresholds, and safeguards.
FRIA: Fundamental rights impact assessment. Analyze who is affected, what decisions the system makes, what risks it creates, and how they are mitigated.
DPIA: Data protection impact assessment (GDPR), mandatory because you process personal data at scale for automated decision-making.
Mandatory human oversight: A person with the competence, authority, and judgment to do so must review the system’s decisions. No candidate may be definitively rejected solely by AI.
Traceability logs: Record every system decision (score, ranking, and reasons) so that any audited decision can be reconstructed.
Fairness metrics: Monitor the system’s differential impact on protected groups (gender, age, origin, and disability).
Registration and conformity: The system must be registered and must have passed a conformity assessment.

No candidate may be definitively rejected solely by AI.

Fairness metrics: how to measure whether your AI discriminates

It is not enough to say that your system is fair. You need to measure it with specific metrics. These are the four fundamental ones:

The 4/5 rule (Disparate Impact Ratio)

The metric most commonly used in regulatory contexts. It compares the selection rate of the least favored group with that of the most favored group. If the ratio is below 80% (4/5), there is evidence of disproportionate impact.

Example: if your AI selects 60% of male candidates and 40% of female candidates, the ratio is 40/60 = 0.67 (67%). It is below 80% → there is disproportionate impact that you must investigate and correct.

Demographic parity

This measures whether the probability of a positive outcome (being selected) is similar across demographic groups. It is the most intuitive metric, but has one limitation: it may conflict with meritocracy if base rates differ between groups.

Equal opportunity (Equalized Odds)

This measures whether the system’s true-positive and error rates are equal across groups. In other words: does the system incorrectly reject qualified candidates at the same rate regardless of their demographic group?

Calibration

This measures whether the system’s scores mean the same thing for every group. If a score of 8/10 for a male candidate predicts the same performance as 8/10 for a female candidate, the system is calibrated.

Important: An impossibility theorem shows that demographic parity and equal opportunity cannot be satisfied simultaneously (unless base rates are identical). You must choose which metric to prioritize according to your context and document that decision.

Mandatory human oversight: how to implement HITL

Human oversight (HITL, Human-in-the-Loop) in recruitment is not a bureaucratic checkbox. It is an operational control that makes the difference between a lawful system and one that exposes you to penalties and lawsuits.

The principles for effective human oversight:

No definitive rejection without human review. AI may prioritize, score, and recommend, but the final decision to reject a candidate must go through a human. Especially in borderline cases.
Supervisor competence. The reviewer cannot be a junior without the necessary judgment. They need to understand what the system does, which variables it uses, where its limits are, and when to correct it.
Real authority to intervene. If the supervisor detects a problem but has no authority to stop or modify the process, oversight is theater. HITL needs veto power.
Confidence thresholds. Configure the system to automatically escalate a case to human review when a candidate’s score falls into the gray zone (neither clearly qualified nor clearly rejectable).
Intervention log. Every human correction must be recorded: who intervened, what they changed, why, and when. This improves the model and provides evidence of compliance.

Responsible implementation: the Gates approach

If you are considering implementing AI in recruitment, follow a progressive implementation model with Gates:

Gate 0: Ideation

Define the value hypothesis: reduce time-to-hire by 30% while maintaining or improving diversity.
Complete the classification form: high risk (Annex III, employment).
Identify the fairness metrics you will monitor.
Appoint the sponsor and the HITL supervisor.

Gate 1: Pilot with real data

Deploy it in a limited recruitment process (one department, one type of role).
Complete the DPIA and FRIA before the pilot.
Monitor the 4/5 rule weekly by gender, age, and origin.
Traffic light: green if 4/5 >80% across all groups. Amber if 70–80%. Red if <70% in any group → KILL.

Gate 2: Production

Expand to all processes with the complete technical documentation.
Dashboard with fairness metrics, HITL intervention rate, and hiring quality.
External bias audit annually or after significant model changes.
Operational withdrawal protocol (kill switch) if the metrics cross critical thresholds.

What a bias audit should include

A bias audit is not a generic report. It is a systematic analysis that must cover:

Training data analysis. Review the representativeness of the historical data used: does it reflect the real diversity of the labor market, or reproduce historical hiring biases? If your company mostly hired one type of profile over the last 10 years, the model will learn to prioritize that profile.
Proxy variable detection. Identify whether the model is using apparently neutral variables that correlate with protected characteristics. ZIP code may correlate with ethnic origin. University type may correlate with socioeconomic status. Writing style may correlate with gender.
Fairness metric calculation. Apply the 4/5 rule, demographic parity, and equalized odds for each protected group: gender, age, origin, and disability. Document the results with numerical data, not subjective assessments.
Stress testing. Expose the system to adversarial scenarios: what happens if you change only the candidate’s name on an otherwise identical CV? Does the score change? If it does, there is bias.
Report and action plan. Document the findings, metrics, identified risks, and corrective measures with deadlines and owners. This report is part of the technical documentation required by the AI Act.

Frequency: Initial audit before Gate 1 (pilot). Quarterly review of fairness metrics in production. Full audit annually, or whenever the model changes, training data is updated, or the system expands to new roles or departments.

AI in recruitment is not optional, but neither is bias

AI in recruitment will become the standard, not the exception. Companies that implement it well will gain a real competitive advantage in speed, cost, and hiring quality. But those that implement it without governance expose themselves to regulatory penalties, discrimination lawsuits, and reputational damage.

The difference is not whether you use AI. It is how you use it: with real human oversight, monitored fairness metrics, documented evidence, and a progressive implementation model that lets you detect and correct problems before they escalate.

If you want to assess whether your current recruitment system carries a bias risk, configure fairness metrics, or prepare the technical documentation required by the AI Act, Impulsa3 conducts bias audits and supports you from Gate 0.

At customer touchpoints, complete this view with the guide to AI transparency in communication.