An AI-based semantic score applied to millions of Dutch professional profiles finds digital job-language concentrated in managerial, professional and ICT occupations but increasingly appearing in hybrid business, marketing and automation roles; career moves into more digital work depend strongly on origin and pathway rather than occurring uniformly across the labour market.

Main Finding

The paper develops an AI-driven, scalable framework to measure occupational digitalisation from large-scale online profile data. It combines embedding-based retrieval and LLM classification to map noisy job-title text to ESCO occupations and introduces a Transparent Digital Semantic Score that captures how semantically “digital” job titles and skills are (relative to a non-digital baseline). Applied to Netherlands profile data, the approach shows digitalisation is uneven: strongest in managerial, professional and ICT occupations but increasingly visible in hybrid business, marketing and automation roles; career moves into digital work are pathway-dependent; and digital capability is multidimensional (technical, hybrid/product, and business-systems/automation).

Key Points

• Method innovation:
  • Hybrid pipeline: dense embeddings (FAISS nearest neighbours) to retrieve top ESCO candidates, followed by an LLM to select the best ESCO code or abstain when uncertain.
  • Digital Semantic Score: an anchor-based, embedding-space index computed as max(similarity to three digital anchor centroids) minus similarity to a non-digital baseline centroid.
• Anchor design: four intentionally compact, interpretable anchor groups:
  • core_digital (e.g., software engineering, AI, cloud, cybersecurity)
  • hybrid_digital (e.g., digital transformation, product owner, e-commerce)
  • business_systems_automation (e.g., ERP, RPA, low-code, SAP)
  • non_digital_baseline (e.g., manual labour, hospitality, construction)
• Data scale and scope:
  • Source: Revelio Labs profile dataset (global >1B profiles), analysis focused on Netherlands job-history subset (LinkedIn-rich market).
  • Time windows: long-run occupational drift compared between baseline 2000–2010 and recent 2020–2024; annual job-title trends also reported.
• Empirical findings:
  • Digital job-title language has increased over time in many ISCO groups but remains concentrated in certain occupations.
  • Hybrid digital vocabulary (marketing, product, automation-related roles) is growing — pointing to skill rebundling rather than only tech-specialist growth.
  • Career-transition analysis: movement toward more digital titles depends on origin occupation; some pathways show larger semantic shifts than others.
  • Skill-level analysis: skills on profiles exhibit multi-dimensional digitality; analyses are profile-level (skills not linked to specific jobs) and should be interpreted accordingly.
• Interpretation: The score is a relative semantic index (not a probability). Subtraction of the non-digital baseline reduces false positives from generic professional language.

Data & Methods

• Data:
  • Revelio Labs profile data with job histories, self-reported job titles and skill lists; focus on Netherlands due to high LinkedIn adoption and DESI ranking.
  • Limitation: skills are profile-level and not tied to specific jobs; some fields are predicted (e.g., gender).
• Occupational mapping:
  • ESCO entries (labels, descriptions, alt/hidden labels) embedded with OpenAI text-embedding-3-large (3,072-d vectors).
  • FAISS used to retrieve top-10 nearest ESCO occupation candidates per job record.
  • A large language model (structured prompt) then selects the best ESCO candidate or abstains if confidence is low.
  • Final occupational codes harmonized to ISCO-08 4-digit level for analyses.
• Digital Semantic Score computation:
  • Job-title and skill text normalized and embedded with text-embedding-3-small; vectors unit-normalized.
  • Centroids computed for each anchor group (average of unit vectors, then normalized).
  • Cosine similarities computed: s_k(t) = cosine(embedding(t), centroid_k).
  • Digital score D(t) = max_k∈{core,hybrid,business} s_k(t) − s_baseline(t).
  • Score interpreted relatively: higher = closer to digital anchors than to non-digital baseline.
• Transition and aggregation methods:
  • Job-to-job transitions built from chronological consecutive records per individual; transition shares row-normalized by source occupation.
  • Occupational digital drift Δ_g = mean recent title score − mean baseline title score (2000–2010 vs 2020–2024).
  • Skill digitalization H_p = average D(s) across parsed skills of person p.
• Robustness / practical choices:
  • Anchor phrases defined manually in English (job titles translated to English).
  • Top-10 retrieval + LLM disambiguation chosen to balance recall and precision; records without valid 4-digit ISCO codes excluded where needed.
  • Results described as descriptive (profile-observed) not population-weighted flows.

Implications for AI Economics

• Measurement and monitoring:
  • Provides a scalable, more semantically rich alternative to keyword lists for tracking occupational digitalisation in near real-time; useful for labour-market observatories and policymakers.
  • Can detect emerging job-title vocabulary and the spread of hybrid digital roles that keyword/occupation-only methods may miss.
• Policy and reskilling:
  • Findings imply targeted reskilling: not all digitalisation is technical — many roles require hybrid or business-systems skills (ERP, RPA, CRM). Training programs and curricula should reflect these multiple digital dimensions.
  • Pathway-dependent mobility suggests mobility supports should be occupation-specific (some source occupations are more likely to transition into digital roles than others).
• Labour-market dynamics and inequality:
  • Uneven digitalisation can create winners and losers; monitoring semantic signals helps identify occupations at risk of mismatch or advantage from digital adoption.
  • The approach can inform shortages forecasts by revealing where digital language and skills concentrate and where they are emerging.
• Methodological implications for research:
  • Demonstrates the utility of combining embeddings + LLMs for mapping noisy multilingual textual labour-market data to harmonized taxonomies (ESCO/ISCO).
  • The anchor-centroid subtraction technique is a transparent way to reduce confounding from generic professional language when measuring semantic traits.
• Cautions and limitations for economic inference:
  • Representativeness: LinkedIn/profile data over-samples certain demographic and occupational groups; not population-representative without reweighting.
  • Skill linkage: skills are not tied to specific jobs, limiting causal inference about skill demands in particular roles or time-specific skill acquisition.
  • Anchor choices, embedding models, translation steps, and provider-predicted fields can introduce biases; results depend on these design choices.
• Research opportunities:
  • Validate semantic scores against vacancy data, surveys, or job-task datasets (O*NET/ESCO task lists).
  • Experiment with alternative anchor sets, multilingual anchors, or unsupervised discovery of digital directions.
  • Link skills to jobs (when possible) for longitudinal causal analysis of skill accumulation and career mobility in digitalising labour markets.
  • Use the pipeline for real-time monitoring of AI adoption impacts, sectoral shocks, and targeted policy evaluation (training subsidies, mobility programs).

Summary: The paper offers a practical, interpretable AI-based toolkit for mapping job-language to standard occupational codes and for quantifying semantic digitalisation across titles and skills. It advances labour-market measurement by capturing nuanced, multidimensional digital trends, but users should account for sample and measurement limitations when drawing policy or causal conclusions.