A nationwide LLM-curated dataset finds that Spanish firms' public reporting of AI use varies sharply by region, industry and size, covering 112,814 firms across 2023 and 2025. The resource classifies whether AI is used internally or embedded in offerings and provides a reproducible baseline for studying diffusion and economic impacts.

Main Finding

The paper provides a reproducible, scalable nationwide dataset that maps explicit, website-disclosed AI adoption for 112,814 Spanish firms in 2023 and 2025 (225,628 firm‑year observations). Using an LLM-based pipeline, it classifies whether firms communicate AI use at all, whether AI is applied internally, and whether AI is embedded in products or services, and aggregates these indicators across provinces (NUTS3), industries, and firm-size categories.

Key Points

• Coverage: 112,814 distinct Spanish firms, two benchmark years (2023 and 2025), producing 225,628 firm‑year observations.
• Indicators: for each firm-year the dataset records (a) presence of AI claims on the website, (b) explicit evidence of AI use in internal processes, (c) explicit evidence of AI embedded in products/services, and (d) whether website content was valid for evaluation.
• Multi-dimensional aggregation: indicators are reported at province–sector–size cell level (NUTS3 × industry × firm-size), enabling territorial, sectoral and size-based comparisons.
• Method: automated pipeline based on large language models that segments website text, semantically filters relevant content, and applies a structured rubric to identify explicit AI evidence.
• Temporal comparability: two benchmark years allow analysis of changes in communicated/adopted AI between 2023 and 2025.
• Reproducibility and scalability: pipeline design enables extension to future years or adaptation to other countries/languages.

Data & Methods

• Data source: firm corporate websites scraped for text content in two snapshots (2023 and 2025).
• Sample construction: 112,814 firms sampled and matched to website content; firm‑year observations total 225,628.
• Processing pipeline:
  • Text segmentation of web pages into coherent chunks.
  • Semantic filtering to surface AI-related candidate passages.
  • LLM-assisted evaluation using a structured rubric that flags explicit mentions/evidence of AI in (i) internal processes and (ii) products/services.
• Output variables:
  • Binary flags per firm-year: AI mentioned (yes/no), AI used internally (yes/no), AI embedded in offerings (yes/no), and content-validity indicator.
  • Aggregated counts and shares by NUTS3 region, industry, and firm-size cell, including the number of valid website observations per cell.
• Quality/limitations (implicit in method):
  • Measures capture explicit, disclosed AI use on public websites — silent/adoptive uses not mentioned online will be missed.
  • Dependent on website coverage and quality; potential selection bias toward firms that maintain informative websites.
  • LLM-based classification introduces risk of false positives/negatives; reproducibility mitigates but does not eliminate model-driven error.
  • Currently language- and web-text–based; non-public AI use (internal tools, proprietary systems) and multimodal evidence (e.g., demos, PDFs, code repositories) may be under‑captured.

Implications for AI Economics

• Measurement advance: provides a fine-grained, reproducible indicator of firms’ communicated AI adoption that complements administrative, patent, and survey measures.
• Spatial and sectoral diffusion: enables analysis of how AI adoption varies across regions (NUTS3), industries, and firm size—useful for mapping hotspots, digital divides, and localized spillovers.
• Temporal dynamics: two-year benchmarks support preliminary assessments of adoption trajectories and short-run diffusion patterns; pipeline can generate further time points for dynamic studies.
• Linking to outcomes: dataset can be merged with firm-level administrative data (employment, productivity, trade, R&D) to study causal relationships between AI adoption and economic outcomes, conditional on disclosure biases.
• Policy targeting: informs regional and sector-specific digitalisation policies (training, infrastructure, subsidies) by identifying low-adoption cells and potential capacity gaps among small/medium firms.
• Methodological template: demonstrates scalable LLM-based measurement that can be adapted cross-country, facilitating comparative studies of AI diffusion and policy effectiveness.
• Research cautions: because the dataset measures publicly disclosed AI, researchers should account for disclosure bias when using it as a proxy for true adoption (e.g., via validation with surveys, admin data, or robustness checks).