Dutch AI in figures: the public sources (CBS, DNB, central government ICT) and how to read them
In the public and business debate about artificial intelligence, revenue expectations, growth percentages and transformation claims fly around your ears. Commercial consultancy reports and vendor presentations often sketch a picture of explosive adoption or, conversely, of acute threat. Anyone who has to take decisions on the basis of facts, however, needs verifiable, independent and periodically recurring data sources. In the Netherlands, publications by official institutions such as Statistics Netherlands (CBS), the Dutch central bank (DNB) and various government registers form the foundation for reliable market analysis.
Reading these public data sources does demand a specific methodological eye. Because of their careful validation processes, public statistics inherently lag behind day-to-day technological reality. Institutions also use delimited definitions that do not always map neatly onto current usage around large language models and generative AI. This article dissects the most important Dutch government and supervisory sources, explains the measurement methods behind them, names the structural limitations and offers a guide to valuing public figures correctly.
1. Statistics Netherlands (CBS): business adoption and ICT use
The most comprehensive source of quantitative data on Dutch business is the annual survey ICT use by businesses from CBS. This survey focuses on organisations with at least ten employed persons and gives a representative picture of the extent to which technologies are taking root across different economic sectors (SBI codes).
CBS measures AI adoption through structured survey forms in which companies are asked about their use of specific AI technologies. Think of speech recognition, natural language processing (NLP), image recognition, machine learning for data use and automated decision-making systems. This breakdown prevents simple automation from being lumped together with advanced models.
The strength of the CBS data lies in the sampling design. By drawing stratified samples based on company size and sector, the bureau can make reliable statements about the Dutch economy as a whole. Anyone who wants to understand how these macro indicators translate into specific market shifts and acquisitions can consult the overview of consolidation among AI start-ups in the Dutch market.
There are, however, clear methodological caveats to the CBS figures:
- Time lag: Data collection usually takes place in the spring, while the definitive report only appears at the end of the calendar year or the beginning of the following year. Figures on 2025 therefore reflect the situation in early 2025.
- Shifting definitions: Until recently, generative AI fell under general categories such as "text analysis" or "machine learning". Only in recent questionnaires are specific questions asked about language models, which means historical trend lines sometimes show a break.
- Self-reporting: Answers are based on the perception of the respondent within the company. An IT manager sometimes interprets "AI use" differently from a board member, which can lead to under- or over-reporting.
2. The Dutch central bank (DNB): financial sector and operational risks
Where CBS looks at the broad economy, the Dutch central bank (DNB) focuses specifically on the financial sector: banks, insurers, pension funds and investment firms. DNB publishes findings through its periodic supervisory reports, thematic studies and the annual risk dashboard.
DNB does not primarily measure the revenue or efficiency gains of AI, but focuses on governance, data quality, explainability and concentration risks. Through mandatory surveys and information requests among supervised institutions, the regulator maps which models are deployed for, among other things, credit assessment, fraud detection, customer acceptance and automated trading.
For a deeper look at how these specific supervisory requirements and risk frameworks work out in practice for financial institutions, see the analysis of AI in the Dutch financial sector.
DNB's data has a character all of its own. It is not a general market statistic, but a risk-oriented measurement. The main points of attention when reading DNB figures are:
| Sector element | DNB measurement method | Point of attention when interpreting |
|---|---|---|
| Concentration risk | Inventory of external cloud and AI suppliers. | Shows dependence on large tech companies, not the quality of the models. | Model governance | Assessment against internal audit frameworks and DORA guidelines. | Reports on process quality, not on the absolute number of active models. |
| Systemic risk | Aggregated stress tests and thematic studies. | The focus is on exception scenarios and outages, not on average daily use. |
3. Central government ICT and the Algorithm Register: transparency within government
For the public sector there are two important sources: the Algorithm Register (algoritmes.overheid.nl) and the central government ICT dashboard. Both fall under the responsibility of the Ministry of the Interior and Kingdom Relations and offer a view of how public bodies manage algorithmic systems and IT projects.
The Algorithm Register is designed as a central database in which ministries, municipalities, water authorities and implementing organisations (such as the employee insurance agency UWV and the tax administration) record which algorithms they use. For each algorithm it describes the purpose, which data are processed, how human intervention is arranged and what the impact on citizens is.
To better understand the legal frameworks, administrative trade-offs and the actual adoption lag among public bodies, read the background on the state of AI in Dutch government.
Although the Algorithm Register is a valuable initiative for democratic scrutiny, the data currently has significant limitations:
- Voluntary nature and coverage: Not all public bodies register their systems at the same pace or with the same level of detail. The register therefore gives a picture of what has been catalogued, not by definition of everything that is running.
- High risks versus everyday tools: The focus is heavily on decision-making algorithms that directly affect citizens. Internal productivity tools, such as AI-driven search systems for civil servants or summarisation modules, are often missing from the registration.
- Absence of quantitative usage figures: The register states that an algorithm exists, but does not broadcast live data on the number of daily transactions or the volumes processed.
4. Energy and data centres: CBS energy statistics and grid capacity
A crucial but often underexposed part of AI in figures is the physical infrastructure. AI models require considerable amounts of power and cooling capacity. CBS publishes detailed monthly and annual data on energy consumption per industry, including the sector "data processing, web hosting and related activities" (SBI 631).
In addition, grid operators such as TenneT and regional grid operators publish data on the capacity of the electricity grid, contracted capacity and the length of the connection queues for data centres.
For the social, spatial and political discussion around this rapidly growing power demand, see the dossier on the debate about data centres and AI computing capacity in the Netherlands.
When analysing energy statistics around AI, analysts should be alert to the following dividing lines:
Electricity measurements at industrial connections make no distinction between a server serving traditional web pages and a GPU cluster training a language model. Figures on "data centre consumption" are therefore always a composite indicator for the whole digital infrastructure, in which AI-specific workloads form only one part.
5. Quality and performance measurement: the gap between adoption and model quality
Public macro statistics from CBS or DNB tell you how many organisations use AI, but say nothing about how well the applied models perform on specific tasks. To make model quality and language ability measurable, open benchmark initiatives and academic evaluation sets have been developed.
Traditional international benchmarks test models mainly on English-language reasoning tasks, medical knowledge or mathematical problems. For the Dutch market, however, performance on Dutch-language legal, administrative and cultural context is decisive. Standardised prompt evaluations, perplexity measurements on Dutch texts and automated NLU test suites fill this gap.
Anyone wanting to investigate how models perform substantively on specific Dutch-language tasks, and which test sets are available for this, can click through to the overview of benchmarks for Dutch-language model output.
When interpreting benchmark data, vigilance is called for against "benchmark contamination" (where test questions have accidentally ended up in the model's training data) and against the fact that a high score on a static test set does not automatically guarantee that a model works flawlessly in a complex business workflow.
6. Labour market and vacancy data: demand for AI expertise
The impact of AI on the Dutch economy can also be read from the labour market. CBS (through the job vacancy survey) and UWV (through analyses of shortage occupations and labour market tightness) monitor how demand for staff develops.
Labour market analysts also use large-scale text mining on public job boards. By searching for keywords such as "machine learning", "prompt engineering", "Python", "LLM" and "data science", the changing demand for specific skills is mapped.
To see where concrete demand for AI talent is building, which roles are growing fastest and how organisations organise their recruitment, see the guide to the AI labour market in the Netherlands.
The main pitfalls with vacancy data are "skill inflation" (the phenomenon whereby employers add AI terms to vacancies to look attractive, while the actual job content barely changes) and the overlap between traditional software development and specific AI engineering.
7. Privacy, GDPR and supervisory reports from the Dutch data protection authority
An indispensable pillar in the Dutch landscape is the Dutch data protection authority (Autoriteit Persoonsgegevens, AP). As supervisor of the processing of personal data and coordinating supervisor for the EU AI Act, the AP periodically publishes market monitors, research reports and enforcement figures.
Among other things, the AP measures how many data protection impact assessments (DPIAs) organisations carry out for AI systems, how many complaints citizens file about automated decision-making, and which sectors show the highest risk of data breaches when using external AI services.
For a practical framework to hold your own organisation against the legal yardstick and privacy requirements, consult the GDPR privacy checklist for Dutch organizations.
It is important to realise that AP reports are inherently "risk-centred". The regulator reports mainly on incidents, insufficient compliance and potential dangers. A rise in the number of notifications to the AP may point to increasing malpractice, but equally to a greater willingness to report and growing awareness within organisations.
8. Four golden rules for interpreting AI statistics critically
To avoid drawing the wrong conclusions from public figures, it is wise to filter every numerical report through four fixed principles:
- Check the sampling frame (N and year): Always look at how many organisations were actually surveyed and in which quarter the measurement took place. A survey of 50 large IT companies says nothing about SMEs in retail.
- Distinguish between pilots and production models: Many surveys ask whether a company is "experimenting with AI". Experimenting in the marketing department is something fundamentally different from a critical business process running entirely on an LLM.
- Watch out for definition changes: If an institution broadens its definition of AI (by adding simple automated rule systems, for instance), the adoption graph shoots up without any new technology actually having been implemented.
- Validate survey data against physical indicators: Compare what companies claim to do (survey) with hard indicators such as IT investments, data centre capacity in use and demand for specialised staff on the labour market.
9. Conclusion: a factual foundation for the Dutch AI transition
Public data sources such as CBS, the Dutch central bank, the Algorithm Register and the Dutch data protection authority offer an indispensable counterweight to commercial claims and market hype. Although public figures lag because of their thorough design and sometimes struggle with shifting definitions, they form the only basis for a verifiable analysis of the Dutch AI market.
As the European AI Act comes further into force and mandatory registrations for high-risk systems mature, the quality and granularity of public AI data in the Netherlands will only increase in the coming years. Anyone who reads the figures with the right methodological lens has the best steering material for strategic decisions.
