An automated LLM-powered pipeline converts unstructured job ads into high-quality labor-market data; Moroccan cybersecurity demand skews to mid-level workers with advanced degrees and CISSP certification.

Main Finding

A fully automated retrieval-augmented generation (RAG) pipeline using large language models can reliably convert unstructured online job ads into structured, analyzable records for cybersecurity labor-market analysis. Applied to 27,360 job ads (1,681 cybersecurity roles) from LinkedIn, Indeed, and Rekrute (Morocco focus), the approach attains very high extraction quality (GPT-4.1 as master evaluator: Correctness 96.276%, Completeness 100%, Relevance 99.952%, Format Adherence 100%, Language Consistency 99.762%, Clarity 99.881%). The processed data show most cybersecurity openings target mid-level candidates (≥3 years’ experience) with advanced degrees (master’s or engineering) and that CISSP is the single most requested certification.

Key Points

• Innovation: A fully automated RAG + LLM framework (GPT-3.5, GPT-4.1, Meta Llama 3) turns unstructured job postings into structured datasets without manual or semi-automated intervention.
• Data: 27,360 job ads collected from LinkedIn, Indeed, and Rekrute; analysis focuses on 1,681 cybersecurity job postings (balanced dataset).
• Model performance: GPT-4.1 used as master evaluator achieves >96% correctness and near-perfect scores on completeness, relevance, format adherence, language consistency, and clarity.
• Labor-market signal: Majority of listings seek mid-level professionals (3+ years), with a preference for advanced degrees (master’s/engineering) and professional certification—CISSP most common.
• Scalability: The automated pipeline addresses prior limitations (manual effort, scalability, NLP expertise requirements) and is potentially applicable across other knowledge-intensive industries.

Data & Methods

• Data sources: Job postings scraped/collected from LinkedIn, Indeed, and Rekrute; national focus on Morocco for sector-specific insights.
• Sample: 27,360 job advertisements total; 1,681 identified as cybersecurity roles used for the sector analysis.
• Pipeline:
  • Retrieval-augmented generation (RAG) approach to combine document retrieval/context with LLM generation for information extraction.
  • LLMs used: GPT-3.5, GPT-4.1, and Meta Llama 3 to parse and structure postings into predefined fields (skills, experience, education, certifications, role level, etc.).
  • GPT-4.1 served as the master evaluator to assess extraction outputs against evaluation criteria.
• Evaluation metrics: Correctness, Completeness, Relevance, Format Adherence, Language Consistency, Clarity. Reported scores (GPT-4.1): Correctness 96.276%, Completeness 100%, Relevance 99.952%, Format Adherence 100%, Language Consistency 99.762%, Clarity 99.881%.
• Outcome variables extracted: role level (experience years), educational requirements, certifications, skills, likely role titles/occupational categories.

Implications for AI Economics

• Labor-demand measurement at low cost and high frequency: Automated extraction enables timely, granular demand-side indicators for cybersecurity skills and credentials, improving responsiveness of economic and labor-market analyses.
• Human capital signaling and returns:
  • Strong employer preference for advanced degrees and CISSP suggests certification and postgraduate credentials carry signaling or productivity value in cybersecurity — implications for private returns to education and certification markets.
  • Policymakers and training providers can use these signals to prioritize program design (e.g., master’s curricula, certification-aligned courses).
• Wage and scarcity dynamics:
  • If demand concentrates on mid-levels with advanced credentials, wage pressure may appear at those experience/credential bands; granular monitoring can detect premium formation or bottlenecks.
• Skill-biased technological change and task composition:
  • The identified skills and certifications help map which tasks are growing and which may be substitutable/augmentable by automation (including AI tools), informing predictions about future labor demand composition.
• Market for labor analytics and decision-making:
  • The demonstrated accuracy and scalability create market opportunities for LLM-based analytics services (firms, public employment services, educational institutions).
  • Lower cost and faster analytics could change how firms recruit and how policymakers allocate reskilling funds.
• Generalizability and scaling effects:
  • Methodology is transferable to other knowledge-intensive industries (finance, healthcare, data science), enabling cross-sector labor-demand comparisons and more efficient reallocation of training resources.
• Limitations / cautions relevant to economic inference:
  • Data coverage & representativeness: Online job ads may not capture informal hiring channels or all firms; geographic focus (Morocco) limits generalizability without replication.
  • Model dependence & bias: Reliance on LLMs introduces risks of extraction errors, bias, and opaque decision rules—need for ongoing validation and auditing.
  • Cost and operational constraints: Running high-performing LLMs (e.g., GPT-4.1) at scale has financial and compute costs; trade-offs between accuracy and cost should be considered.
  • Causal inference limits: Extracted posting content is descriptive of posted requirements, not necessarily of realized hires, wages, or productivity—complementary data (wages, placement outcomes) are needed for causal economic analysis.

Concluding note: This study shows LLM-driven RAG pipelines can produce high-quality, scalable labor-market datasets for cybersecurity and similar sectors, enabling more timely economic measurement and policy-relevant insights while inviting attention to representativeness, model transparency, and cost considerations.