MAP-AI: Firm-Level AI Adoption, Roles, and Technology Types in Europe

Aggregated indicators of firm-level artificial-intelligence adoption, AI ecosystem roles, and AI technology types for four European countries, derived from the MAP-AI (Mapping Artificial Intelligence in Firms) dataset and described in:

Garbers, Julio and Terry Gregory. "The Diffusion of Artificial Intelligence Across Firms: Evidence from Europe." IZA Discussion Paper, No. 18434 (2026).

MAP-AI uses a Large Language Model to read firm website content and identify whether a firm adopts AI (uses, develops, or deploys AI technologies), what role it plays in the AI ecosystem, and which types of AI technology it works with. The indicator captures realized AI adoption rather than potential AI exposure. See the paper for the full classification methodology and validation.

What is in this repository

This repository publishes aggregated firm-share tables. They cover the full matrix of two aggregation geographies × three indicator families:

	AI involvement	AI ecosystem roles	AI technology types
Industry × country	ai_adoption_share_by_industry_country	ai_role_share_by_industry_country	ai_type_share_by_industry_country
NUTS region	ai_adoption_share_by_nuts_region	ai_role_share_by_nuts_region	ai_type_share_by_nuts_region

Each table is provided as both .parquet and .csv. All tables are in long format and share the same firm population and denominators (see Methodology).

File (stem)	Rows	Unit
ai_adoption_share_by_industry_country	60,584	classification × code × digit level × country × year
ai_role_share_by_industry_country	484,672	… × role (8)
ai_type_share_by_industry_country	666,424	… × technology type (11)
ai_adoption_share_by_nuts_region	5,922	NUTS code × level × year
ai_role_share_by_nuts_region	47,376	… × role (8)
ai_type_share_by_nuts_region	65,142	… × technology type (11)

Coverage

Countries: Belgium, France, Germany, Luxembourg (plus a Pooled aggregate that combines all four, in the industry × country tables only). The region tables need no pooled aggregate: NUTS regions are nested within a single country, so NUTS level 0 already provides the per-country totals.

Years: 2016–2024 (9 years).

Industry codes (NACE Rev. 2 and NAICS 2022, at 1–4 digit levels):

Classification	1-digit	2-digit	3-digit	4-digit
NACE Rev. 2	21 sections (A–U)	88 divisions	272 groups	615 classes
NAICS 2022	9 first-digit groups	24 subsectors	96 industry groups	304 industries

NACE 1-digit codes are the official section letters (A–U); all higher NACE levels and all NAICS levels are numeric. Note: NAICS has no official 1-digit level (its sectors are 2-digit, some spanning ranges such as 31–33). The NAICS digit_level = 1 rows are a non-standard grouping by the first digit of the 4-digit code, retained for backward compatibility; use the 2-digit level for sector-level analysis.

NUTS regions (Eurostat NUTS 2024 classification). A region appears as a row once at least one firm is located in it; exactly as for industries, its shares are reported only when the relevant denominator reaches 30 (n_firms ≥ 30 for ai_adoption_share). In practice every region present clears the 30-firm threshold, so no regional adoption share is suppressed.

NUTS level	Meaning	Code length	Regions present
0	Country	2	4
1	Major socio-economic regions	3	34
2	Basic regions	4	77
3	Small regions	5	543

Schema

AI-involvement tables

ai_adoption_share_by_industry_country and ai_adoption_share_by_nuts_region:

Column	Type	Description
classification	string	"nace" or "naics" (industry table only)
code	string	Industry code, e.g. "C", "62", "6201" (industry table only)
digit_level	int	Industry aggregation level: 1, 2, 3 or 4 (industry table only)
nuts_code	string	NUTS region code, e.g. "FR", "FR1", "FRI11" (region table only)
nuts_level	int	NUTS aggregation level: 0, 1, 2 or 3 (region table only)
country	string	"Germany", "France", "Luxembourg", "Belgium" (and "Pooled" for the industry table)
year	int	Year of observation (2016–2024)
ai_adoption_share	float	Share of classified firms that are AI-active (null if n_firms < 30)
n_firms	int	Number of classified firms in the cell

Role and technology-type tables

The four ai_role_* / ai_type_* tables add one row per category. They share this schema (replace role→type, ai_role→ai_type for the type tables, and the geography columns as above):

Column	Type	Description
(geography columns)		as in the involvement tables
year	int	Year of observation (2016–2024)
ai_role / ai_type	int	Category code (roles 1–8, types 1–11; see Taxonomies)
ai_role_label / ai_type_label	string	Human-readable category label
role_share_among_ai_firms / type_share_among_ai_firms	float	Share of AI-active firms in the cell that fall in this category (null if n_ai_firms < 30)
role_share_among_all_firms / type_share_among_all_firms	float	Share of all classified firms in the cell that fall in this category (null if n_firms < 30)
n_ai_firms	int	Number of AI-active firms in the cell (denominator of *_share_among_ai_firms)
n_firms	int	Number of classified firms in the cell (denominator of *_share_among_all_firms)

Methodology

Firm population and unit

The unit of observation is the firm-year. A firm contributes to a year if and only if it could be matched to classified website content and that year falls in 2016–2024. n_firms therefore counts classified firms, and is identical across the involvement, role, and type tables for the same cell.

Cumulative indicators

All indicators are cumulative: a firm counts as AI-active / as having a given role / using a given technology type in year y if it was ever classified as such up to and including year y. This is the measure underlying the paper's headline diffusion series.

ai_adoption_share

The share of classified firms in the cell that are AI-active — the mean of the binary cumulative AI-adoption indicator over the firm-years for which it is defined. The industry-country adoption table reproduces the previously published table exactly.

A small number of classified firm-years have an undetermined adoption indicator. The mean drops them from its denominator, whereas n_firms still counts them, so ai_adoption_share can differ from n_ai_firms / n_firms (the role/type tables' all-firm denominator) by at most about 0.003, in a minority of cells. Treat ai_adoption_share as the authoritative adoption measure (it is the one used in the paper).

Roles and technology types: two denominators

A firm is assigned roles and technology types only if it is AI-active, and it can have several at once (the categories are not mutually exclusive — within a cell the category shares do not sum to one). To make both natural readings available, each category row reports two shares:

• *_share_among_ai_firms = firms in the category ÷ n_ai_firms — the composition of the local AI ecosystem (the framing used in the paper).
• *_share_among_all_firms = firms in the category ÷ n_firms — the share of the whole local firm population, on the same n_firms denominator that is reported alongside it.

Both denominators (n_ai_firms and n_firms) are reported, so for any cell above its suppression threshold the underlying category firm count is recoverable as round(share_among_all_firms × n_firms) or, equivalently, round(share_among_ai_firms × n_ai_firms).

Category structure

• Roles — categories 1–6 are substantive ecosystem roles, 7 is Other AI role, and 8 is AI role cannot be inferred (7 and 8 are mutually exclusive fallbacks).
• Technology types — categories 1–9 are substantive technology types, 10 is Other AI technology type, and 11 is AI type cannot be inferred (10 and 11 are mutually exclusive fallbacks).

Minimum-firm suppression

To avoid reporting unreliable shares from small samples, a share is set to null when its denominator is below 30: ai_adoption_share and *_share_among_all_firms when n_firms < 30, and *_share_among_ai_firms when n_ai_firms < 30. Counts (n_firms, n_ai_firms) are always reported.

Code validation

Industry codes are validated against the official classification files (Eurostat NACE Rev. 2 RDF and the 2022 NAICS Structure); NUTS codes against the Eurostat NUTS 2024 geometry file. Higher aggregation levels are derived from the validated finest level (4-digit industries; 5-character NUTS3 regions).

Taxonomies

AI ecosystem roles

Code	Label
1	AI core technology provider
2	AI application and solution developer
3	AI adopter or innovative user
4	AI research or support organization
5	AI investor or incubator
6	AI data infrastructure provider
7	Other AI role
8	AI role cannot be inferred

AI technology types

Code	Label
1	Generative AI & conversational systems
2	Text & language (non-generative)
3	Image & video (non-generative)
4	Speech, audio & music
5	Predictive & analytical applications
6	Recommendation & personalization
7	Autonomous systems & robotics
8	Security & fraud detection
9	Foundational AI/ML technologies
10	Other AI technology type
11	AI type cannot be inferred

Main data source

Common Crawl

Citation

If you use this data, please cite:

@techreport{garbers_and_gregory_diffusion_ai,
  title       = {The Diffusion of Artificial Intelligence Across Firms: Evidence from Europe},
  author      = {Garbers, Julio and Gregory, Terry},
  year        = {2026},
  institution = {Institute of Labor Economics (IZA)},
  type        = {IZA Discussion Paper},
  number      = {18434},
  url         = {https://docs.iza.org/dp18434.pdf}
}

License

This dataset is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt this data for any purpose, provided you give appropriate credit by citing the paper above.