The Smart Labour Market Oversight Model is an AI-powered system developed by Saudi Arabia’s HRSD to proactively detect fraud and non-compliance in the national labour market. It analyses more than 90 million transactions annually to identify suspicious patterns, including ghost employment and occupational inconsistencies, across a labour market comprising more than 5 million establishments and 11 million workers.
Initiative overview
Saudi Arabia’s labour market includes more than 5 million establishments and 11 million workers and processes over 90 million work-permit transactions annually. Before the Smart Labour Market Oversight Model, regulatory enforcement relied on reactive, complaint-driven mechanisms.
Inspectors manually prioritised cases without analytical support, leading to inconsistent decisions, slow response times averaging 25 days per cycle, and limited capacity to detect sophisticated fraud patterns. These included ghost Saudisation, where businesses falsely register Saudi nationals to meet Nitaqat quotas, and inconsistencies between workers’ registered occupations and their actual activities.
Objectives
The Smart Labour Market Oversight Model was developed to:
- transform enforcement from a reactive, complaint-based process into a proactive, data-driven system;
- reduce subjective variation in regulatory decision-making through standardised, AI-supported risk-scoring models;
- substantially reduce the operational cost and time required to identify and act on potentially non-compliant establishments; and
- protect the rights of national and migrant workers by supporting more consistent and objective regulatory treatment across establishments.
How it works
The system combines multiple AI technologies in a unified suspicion-scoring engine. Large Language Models, specifically Google Gemini, analyse the semantic compatibility between workers’ registered occupations and establishments’ actual economic activities, producing a five-tier classification.
Natural Language Processing extracts behavioural patterns from unstructured Arabic-language data. Classification and regression models process millions of records to rank potentially suspicious cases by risk level.
All outputs are integrated through a Semantic Layer that unifies data from six government platforms, including Ajeer, Qiwa and GOSI, through ETL and ELT pipelines. This supports real-time data flows without manual intervention. MLOps and AIOps pipelines enable continuous model retraining as fraud patterns evolve.
Other relevant details
Results, outcomes and impacts: Reported quantitative results include: an 88% automation rate in suspicion-detection operations; a 90% reduction in operational costs, reportedly saving SAR 2.25 million per month; more than 40,000 work hours freed each month; a reduction in the decision cycle from 25 days to three days; quality review of more than 100,000 work permits; correction of more than 1,000 establishments without field intervention; and active coverage of more than 3 million establishments and 10 million workers. Results were measured through internal HRSD operational dashboards, monthly KPI reporting to executive committees, and benchmarking against national competitiveness indices.
Challenges and lessons learned: Data complexity: Integrating more than 90 million transactions from six government platforms required a Semantic Layer to resolve definitional conflicts and harmonise data schemas. AI calibration: Fine-tuning LLMs for the Arabic labour-market context required extensive prompt engineering and few-shot learning before reliable performance could be achieved. Evolving fraud: Manipulation strategies change continuously, requiring ongoing investment in MLOps to keep models current. Key lessons include the value of weekly operational reviews in supporting rapid iteration and the importance of embedding governance frameworks, including ISO 27001 and ISO 38500, from the outset. The initiative also demonstrates that a robust Semantic Layer should precede AI development, as data unification is fundamental to reliable outputs.
Future evolution: The system is designed for long-term scalability. Planned expansions include the detection of additional fraud typologies beyond ghost Saudisation and occupational inconsistency, integration with additional national registries, development of a predictive risk-scoring dashboard for policy simulation, and potential sharing of the model framework with Gulf Cooperation Council partner countries as a regional compliance standard. The cloud-native, modular architecture allows the system to scale with increasing data volumes without requiring major system reconstruction.
