Leveraging Machine Learning for Inclusive Employment (Case Study: Italy)

In the pursuit of the UN SDGs, SDG 8 calls for sustained, inclusive economic growth and decent work for all, including persons with disabilities, who remain significantly underrepresented in labor markets worldwide. While many countries have legislative quotas and employment protections, structural barriers, administrative bottlenecks, and inefficient job-matching systems often prevent meaningful inclusion.

A production-ready machine learning system developed in Italy demonstrates how data science can directly address this gap. By combining advanced predictive modeling, participatory requirements engineering, and scalable system architecture, the platform transforms disability employment services from a manual, time-intensive process into a high-precision, data-driven matching system. This case illustrates how artificial intelligence can serve SDG 8 by improving access to decent work for vulnerable populations while increasing institutional efficiency.

Context and Challenge

Despite legal frameworks and quota-based hiring mandates designed to support disability employment, labor market participation among people with disabilities in Italy remains disproportionately low. Nationally, only 3.5% of people with disabilities are employed, highlighting a persistent gap between policy intent and actual inclusion.

Part of this challenge stems from operational constraints within employment services. Employment centers traditionally manage disability job matching through manual review processes. Caseworkers evaluate candidate profiles, analyze job descriptions, consider geographic feasibility, and assess readiness for work. Each case may require 30–60 minutes of evaluation time. As caseloads increase, this creates severe operational constraints, such as: limited throughput of candidates, inconsistent evaluation standards across caseworkers, delays in connecting individuals to suitable vacancies, and risk of mismatched placements.

These inefficiencies do not merely affect administrative performance since they directly impact livelihoods. Delayed or inaccurate job matching reduces employment opportunities for individuals who already face systemic disadvantage.

The challenge, therefore, was twofold: increase the scale and speed of matching while preserving fairness, interpretability, and human oversight in a socially sensitive domain.

Data-Driven Solution

To address these barriers, researchers developed a production-ready machine learning system for inclusive employment, designed in collaboration with employment service professionals (Centro per l’Impiego di Villafranca di Verona), who co-designed system requirements and validated outputs. The system applies ensemble learning techniques and scalable architecture to automate and enhance disability job matching.

Methodology

The system’s architecture integrates multiple machine learning models operating in parallel. Its methodology includes:

1. Data Encoding and Feature Engineering: Candidate profiles and job descriptions are transformed into structured, machine-readable formats. Features include semantic similarity between skills and job requirements, geographic proximity, employment readiness indicators, and compatibility metrics.
2. Ensemble Modeling and Optimization: Seven machine learning models are combined into an ensemble framework. Hyperparameter tuning was conducted using automated optimization tools, improving predictive accuracy and robustness.
3. Multi-Dimensional Scoring Engine: The platform calculates composite scores for each candidate (job pair & ranking opportunities) based on compatibility across multiple criteria rather than relying on a single matching rule.
4. Performance Engineering for Scalability: The system was engineered to handle high-volume evaluations efficiently. It can process 500,000 candidate–company match combinations in under 10 minutes, with sub-100 millisecond response times per individual query.
5. Human-in-the-Loop Oversight: Crucially, AI does not replace employment professionals. Instead, it generates ranked recommendations, while final decisions remain under human supervision. This preserves contextual judgment and ensures ethical accountability.

Implementation and Results

The system was deployed in collaboration with employment center professionals and evaluated under real operational conditions. The quantitative results demonstrate both technical effectiveness and practical value:

• 1% F1-score accuracy in predicting appropriate candidate–job matches
• Processing capacity of 500,000 match combinations in under 10 minutes
• Sub-100 ms response time for individual match evaluations
• Expert validation indicating 60–100% increase in employment center throughput

These results represent a substantial improvement over traditional manual processes. Caseworkers reported that AI-generated rankings significantly reduced time spent on preliminary screening, allowing them to focus on counselling and employer engagement.

Importantly, the system maintained interpretability and transparency in recommendations, both are critical factors in disability employment services, where trust and fairness are paramount.

Outcomes and Impact

The platform directly advances SDG 8 by improving access to decent work for people with disabilities while strengthening institutional efficiency.

• Increased Access to Employment: Higher matching accuracy improves the likelihood of successful placements and long-term retention.
• Operational Efficiency: Doubling service throughput allows employment centers to support more individuals without proportional increases in staffing.
• Scalability: The production-ready architecture enables replication in other regions or national systems.
• Ethical AI Deployment: By embedding human oversight and participatory design, the system demonstrates responsible AI implementation in a sensitive social domain.

The broader implication is structural: employment services can transition from reactive, manual evaluation systems to proactive, data-driven inclusion platforms. This shift enhances both social equity and economic productivity.

Conclusion

The AI-driven disability job matching platform in Italy exemplifies how machine learning can serve SDG 8: Decent Work and Economic Growth. By combining ensemble modeling, scalable system design, and human-centered governance, the system achieves 90.1% predictive accuracy, processes half a million match combinations in minutes, and increases institutional capacity by up to 100%.

This case demonstrates that inclusive employment is not only a policy ambition but a technical possibility. When governments and researchers collaborate to design AI systems grounded in social responsibility, data becomes a powerful instrument for expanding access to decent work.

As countries seek to build more inclusive labor markets, this initiative provides a replicable blueprint: harness data, preserve human judgment, and deploy AI not as a replacement for people but as a multiplier of opportunity.