Two in three generative AI applications that Dutch public bodies list in the national Algoritmeregister depend on a US company. Lindstead analysed all 1,574 entries from 341 organisations in the register's export of 4 October 2026. 80 of the 124 generative AI entries (64.5 percent) name a US-controlled supplier or state that the data is processed on a US cloud.1
Summary
- Two in three generative AI entries have a US link. 80 of the 124 entries that describe generative AI (64.5 percent) name a US-controlled supplier or say the data runs on a US cloud. TNO's monitor of December 2025 found 44 of 81 applications (54 percent) on US foundation models.
- Dutch supplier names hide US clouds. In 114 entries the supplier is not American, but the entry's own text says the data passes through a US cloud or AI service. Of 126 anonymisation entries from the Dutch firms xxllnc and DataMask, 77 state that document text is analysed on a Microsoft Azure server.
- More than half of the organisations have a US link. 185 of the 341 organisations in the register have at least one entry with a US supplier or US cloud processing, including 133 of the 208 municipalities. Across all 1,574 entries the share is 17.8 percent.
How many government AI systems depend on US companies?
An entry has a US link when it names a US-controlled supplier, or when its own text says the data is processed on Microsoft Azure, OpenAI, Google Cloud or another US cloud or AI service. Across the whole register, 280 of 1,574 entries (17.8 percent) meet that test.
The register has no reliable field that marks a system as AI. Lindstead used keyword proxies instead: 570 entries describe AI, and 124 describe generative AI such as language models, GPT or Copilot. The closer an entry gets to generative AI, the more often it has a US link.
Generative AI leans on US firms most
64.5 percent of generative AI entries have a US link, against 17.8 percent of all entries.
Two in three generative AI entries in the Dutch register have a US supplier or run on a US cloud.
Register entries with a US link, by type (number of entries), %
¹ Keyword proxy based on the entry text; see the method section. ² Impactvolle algoritmes, the register's category for algorithms with a direct effect on citizens or businesses. US link: a US-controlled supplier is named, or the entry's own text states processing on a US cloud or AI service.
Source: Algoritmeregister export of 4 October 2026; Lindstead analysis
View data table
| Category | Value | Group |
|---|---|---|
| Generative AI¹ (124) | 64.5% | 2 |
| AI systems¹ (570) | 32.3% | 2 |
| All entries (1,574) | 17.8% | 1 |
| High-risk AI (41) | 14.6% | 1 |
| Impactful² (709) | 4.4% | 1 |
Group 1: Register category. Group 2: Lindstead keyword proxy. Unit: %.
Of the 80 generative AI entries with a US link, 48 name a US supplier. The other 32 name a Dutch supplier, or none, and describe US processing in their own text. Counting only systems in use, the share is 64 of 100. The 80 entries come from 59 organisations.
These figures update TNO's government-wide monitor of December 2025. TNO counted 81 generative AI applications, and found that 44 of them (54 percent) use foundation models from US companies.2 The register count is broader: it covers US suppliers and US clouds as well as models.
What sits behind Dutch supplier names?
The supplier field shows who sells the software. It does not show where the data is processed. In 114 entries the named supplier is not American, but the entry's own text says the data passes through a US cloud or AI service. For 105 of them the supplier is Dutch. These entries come from 103 organisations.
Anonymisation software is the largest group. Municipalities use it to black out personal data before they publish documents. Amstelveen, for example, names Woo (freedom of information) requests and permit decisions. The Dutch firm xxllnc sells this software, partly under the DataMask brand it bought in December 2023. Of the 126 xxllnc and DataMask anonymisation entries, 77 state that document text is analysed on a Microsoft Azure server. Most add that the content is not stored there. Computable reported in January 2026 that xxllnc moved parts of the anonymisation package from the Dutch host TransIP to Amazon Web Services.3 Both point to US cloud providers.
Tolkie, Struck.build and Weglot follow, with five such entries each. A register that lists only the supplier therefore understates US processing. The publication standard has a supplier field, but no field for the hosting provider or the underlying model.4
Which US companies supply the most government algorithms?
18 of the 207 distinct suppliers named in the register are US-controlled. Together they appear in 151 entries. The largest is a Dutch brand. Kiteworks of San Mateo, California, announced its acquisition of secure-mail service Zivver in June 2025.5 Binnenlands Bestuur reported the deal at the time, and the cabinet answered parliamentary questions about it on 16 December 2025.6,7 Neither gave a count of government users. The register shows 73 entries at 73 organisations.
US-controlled suppliers in the Dutch algorithm register, by number of entries
| Supplier | Parent company | Based in | Entries | Ownership note |
|---|---|---|---|---|
| Zivver | Kiteworks | San Mateo, CA | 73 | Announced June 2025 |
| Microsoft | Microsoft Corporation | Redmond, WA | 40 | |
| ZyLAB | Reveal Data Corp. | Chicago, IL | 15 | IPRO 2021, Reveal 2023 |
| OpenAI | OpenAI | San Francisco, CA | 7 | |
| UiPath | UiPath Inc. | New York, NY | 6 | |
| Oracle | Oracle Corporation | Austin, TX | 2 | |
| GitHub | Microsoft Corporation | Redmond, WA | 2 | Microsoft since 2018 |
| Avolve Software | Avolve Software | Phoenix, AZ | 2 | |
| Ten others¹ | Google, Salesforce, IBM and more | US | 10 | One entry each |
Source: Algoritmeregister export of 4 October 2026; owners from company press releases, websites and public company records, checked 4 October 2026. ¹ Qlik, Teledyne FLIR, ATLAS.ti (Lumivero), Google, Bloomberg, Orbis (Moody's), Algolia, Salesforce, IBM and OneTrust. An entry that names two US suppliers counts for each, so the column adds up to more than 151.
With 73 entries, Zivver is now the largest US-controlled supplier in the register, ahead of Microsoft with 40. 44 of the 73 Zivver entries were first published before the takeover, and the register shows no sign of the change. ZyLAB, an e-discovery tool with 15 entries, has a similar history: US firm IPRO bought it in 2021, and Reveal of Chicago took over IPRO in 2023.8
Ownership changes move the numbers. Counting Zivver as Dutch, as it was before June 2025, the US share of entries that name a supplier falls from 16.1 to 8.3 percent.
How many organisations have an algorithm with a US link?
185 of the 341 organisations in the register (54 percent) have at least one entry with a US link. Among municipalities it is 133 of 208. Without Zivver, 207 entries at 146 organisations (43 percent) remain.
Entries with a US link by type of organisation
| Type of organisation | Entries | With a US link | Share |
|---|---|---|---|
| Municipalities | 988 | 209 | 21.2% |
| National government (Rijk) | 303 | 9 | 3.0% |
| Independent bodies (ZBO) | 80 | 12 | 15.0% |
| Joint municipal bodies | 71 | 26 | 36.6% |
| Provinces | 63 | 14 | 22.2% |
| Water boards | 25 | 9 | 36.0% |
| Other | 44 | 1 | 2.3% |
| All organisations | 1,574 | 280 | 17.8% |
Source: Algoritmeregister export of 4 October 2026; Lindstead analysis. Type from the register's organisation code, with a manual list for 19 numeric codes. Other: foundations, probation services and similar bodies.
Joint municipal bodies and water boards show the highest shares, 36.6 and 36.0 percent, from small numbers of entries. The national government sits at the other end with 9 of 303. A section below explains why that figure is low.
Is the share of US technology growing?
Yes. Newer algorithms more often have a US link. Of the algorithms that went into use in 2017 or earlier, 5 of 199 (2.5 percent) have one. Of those that started in 2025, 57 of 171 (33.3 percent) do.
New algorithms run on US technology more often
One in three algorithms that went into use in 2025 has a US link.
The share of government algorithms with a US link rose from 2.5 to 33.3 percent by start year.
Register entries with a US link, by year the algorithm went into use, %
¹ 1 January to 4 October 2026. Year the organisation states the algorithm went into use. 1,146 of 1,574 entries give a readable year. Entries per period: 199, 149, 226, 184, 163, 171 and 54. Algorithms taken out of use before the snapshot are partly missing.
Source: Algoritmeregister export of 4 October 2026; Lindstead analysis
View data table
| Category | Value |
|---|---|
| 2017 or earlier | 2.5% |
| 2018 to 2020 | 16.1% |
| 2021 to 2022 | 10.6% |
| 2023 | 22.8% |
| 2024 | 24.5% |
| 2025 | 33.3% |
| 2026¹ | 31.5% |
Value: Entries with a US link. Unit: %
Generative AI drives much of the rise. 49 of the 171 entries that started in 2025 describe generative AI, against 4 of 184 in 2023. The year is the start date the organisation entered. Algorithms switched off before the snapshot are partly missing, so the early years may understate what was in use at the time.
Do US companies run algorithms that decide about citizens?
Mostly not. The US link sits mainly in office tools: secure e-mail, anonymisation, document search and copilots. 185 of the 280 US-linked entries fall in the register category for internal operations. Of the 709 impactful algorithms, 31 (4.4 percent) have a US link. Of the 41 high-risk AI systems, 6 do.
The exceptions matter, because they touch case files, income support and jobseekers.
Impactful and high-risk entries with a US link: five examples
| Organisation | What the system does | US technology | Category |
|---|---|---|---|
| Instituut Mijnbouwschade Groningen | Checks NAM damage files | GPT-4 in its own Azure | Impactful |
| Gemeente Delft | Matches jobseekers to jobs | GPT-4o on Azure (NextHuman) | High-risk AI |
| Kamer van Koophandel | Classifies business activities | ChatGPT 4o; Azure OpenAI | Impactful |
| Gemeente Arnhem | Grants two income supplements | Microsoft Power BI | Impactful |
| Gemeente Amsterdam | Answers website questions | Salesforce Einstein chatbot | Impactful |
Source: entry texts in the Algoritmeregister, export of 4 October 2026; names link to the entry. Kamer van Koophandel (ChatGPT 4o for its SBI classifier, Azure OpenAI for its input assistant) and Gemeente Arnhem (two supplements) each have two entries. NextHuman is a Dutch supplier; the entry says the data sits in its Azure environment within the EU.
In Groningen and Delft a language model reads files about individual people. In Arnhem a Microsoft tool supports the allocation of two income supplements. Keeping data inside the EU limits where it sits. It does not remove the legal reach of a US provider, as the CLOUD Act section explains.
How far does Microsoft reach into government algorithms?
42 entries at 36 organisations name Microsoft or its subsidiary GitHub as supplier. Microsoft's role is larger than that count. Azure appears as the processing platform in 132 entries at 112 organisations; in 100 of those the named supplier is Dutch.
30 entries at 29 organisations register a Microsoft Copilot deployment. Three organisations classify Copilot as a high-risk AI system: Gemeente Zeist, Gemeente Bergen (L) and the Commissariaat voor de Media. Seven entries list it as impactful and 20 as another algorithm. The AI Act's high-risk list names uses, such as decisions on benefits. A general tool like Copilot can therefore land in different categories. The AI Act guide for the public sector sets out which uses are high-risk.
Why does the national government barely show up?
Ministries and national agencies account for 303 entries from 47 organisations. Only 46 of those entries name a company, and 9 have a US link. The low count shows what ministries publish. Other sources show what they use.
The Algemene Rekenkamer found in January 2025 that more than half of the Rijk's most important public cloud services are bought from Amazon, Microsoft and Google. For 84 of the 126 most important services (67 percent), no risk assessment had been made.9 In October 2024 it counted 433 AI systems at 70 government organisations, and found 5 percent of them in the register.10 The Belastingdienst decided in December 2025 to stay on Microsoft 365, citing the lack of a European alternative.11
Municipalities account for 988 of the 1,574 entries. The register shows municipal practice in detail, and the Rijk's only in outline.
What does the US CLOUD Act mean for these systems?
The CLOUD Act lets US authorities order a provider under US jurisdiction to hand over data in its possession, custody or control, wherever that data is stored.12 An Azure data centre in the Netherlands is still run by a US company. The same now holds for Zivver. Encryption can limit what a provider is able to hand over, and Zivver advertises zero-knowledge encryption.5 Ownership still decides which law the provider answers to.
The Rijk's revised cloud policy of 3 July 2026 responds to this risk. E-mail and documents may only go into the public cloud after a risk analysis and exit plan, and with the approval of the minister responsible. Storage and processing must take place in the EEA or Switzerland. Organisations register material public cloud use and report it to the CIO Rijk each year. The policy also notes the Tweede Kamer's wish that by 2029 at least 30 percent of the Rijk's cloud services come from Dutch or European soil.13
The policy binds ministries and is expected of independent administrative bodies. Other public bodies, including municipalities, are advised to follow it. Its rule on e-mail and documents implements a motion adopted on 18 March 2025.14 The guide to the CLOUD Act and AI data sets out what the law reaches for prompts, embeddings and logs.
What should public bodies do with these findings?
The register shows that the data path often differs from the supplier name. Lindstead recommends five steps:
- Record the hosting provider and the AI model for every entry, next to the supplier.
- Ask each supplier which cloud and AI services process the data, and who owns them.
- Check the ownership of key suppliers once a year. Zivver changed hands in 2025; its entries do not show it.
- For impactful and high-risk systems, record the jurisdiction in the DPIA and, from December 2027, in the fundamental rights impact assessment.
- Where a system reads files about people, test whether a model on European or own infrastructure does the job.
The public sector page covers use cases and the rules that apply. Lindstead deploys AI on infrastructure the organisation controls, and maps AI Act, GDPR and NIS2 obligations to each AI system.
How was this measured, and what are the limits?
Data. Lindstead downloaded the register's full export on 4 October 2026: 1,574 entries from 341 organisations.1 Every figure on this page comes from one script run on that file. The register changes daily, so later counts will differ.
Supplier ownership. 394 different spellings in the supplier field were mapped to 207 suppliers. For each, Lindstead recorded the parent company and the headquarters country of the top operating company. Dutch subsidiaries of US companies count as US. An entry counts as US-controlled if any named supplier is. Ownership is as of October 2026.
US processing in the text. A search for US cloud and AI products, such as Azure, Microsoft 365, Copilot, OpenAI, GPT, Google Cloud, AWS and Anthropic, matched 217 entries. Each was read by hand. 174 describe processing on a US service, 26 mention one only as a tool, and 17 were false matches.
AI proxies. An entry counts as AI when its text uses terms such as machine learning, AI, neural network, language model or computer vision, after negations such as "geen AI" are removed. All 41 high-risk entries count as AI. Generative AI covers terms such as language model, LLM, generative, GPT, ChatGPT, Copilot, Gemini, Claude and Mistral.
Limitations
- The register is voluntary and incomplete. The Ministry of the Interior wrote in September 2025 that a legal obligation is not yet appropriate.15 The Rekenkamer found 5 percent of the Rijk's AI systems in it.10
- Office tools are over-represented. The Autoriteit Persoonsgegevens found that almost 20 percent of registrations between June and December 2025 were secure-mail or anonymisation tools.16
- Entries are not systems. Municipalities copy templates: 70 xxllnc entries share one sentence about Azure, and the 73 Zivver entries are 73 organisations registering one product.
- The supplier field is free text, and empty or "n.v.t." in 355 entries (22.6 percent). US hosting counts only when an entry describes it, so every US figure on this page is a floor.
- Every supplier classed as US was checked against a primary or secondary source. 80 suppliers with 117 entries could not be verified and are classed as non-US, so errors there can only raise the US figures.
- Borderline cases count as non-US: Canva (Australian operations, Delaware parent, 2 entries) and Idemia (French, US private-equity owner, 2 entries). Counting both as US adds at most 4 entries.
- The start year is free text; 1,146 entries give a readable year.
Earlier work: TNO mapped generative AI applications and their foundation models for the Ministry of the Interior.2 Algorithm Audit compared AI inventories with the register in August 2025.17 This analysis adds supplier ownership, US processing behind non-US suppliers and coverage of the whole register.
Download the data
Download the data (CSV): one row per register entry, 1,574 rows. Each row has the organisation and its type, the entry name, status and categories, start year, suppliers and their countries, the US supplier and US processing flags, the AI proxies and a link to the entry in the register. Names and categories stay in Dutch, as in the register.
Cite: Lindstead (2026), Dutch government algorithms and US tech: register analysis, Algoritmeregister export of 4 October 2026. CC BY 4.0 applies to Lindstead's analysis; the register's own terms apply to the source data. Research questions and corrections.
Frequently asked questions
-
In the Algoritmeregister export of 4 October 2026, 280 of 1,574 entries (17.8 percent) name a US-controlled supplier or say their data is processed on a US cloud or AI service. They come from 185 of the 341 organisations in the register. Many entries do not describe their hosting, so this is a floor.
-
80 of the 124 register entries that describe generative AI (64.5 percent) have a US-controlled supplier or state that data is processed on a US cloud. Generative AI is a keyword proxy: entries that mention language models, GPT, Copilot or similar terms.
-
Zivver is a Dutch secure-mail brand. Kiteworks, based in San Mateo, California, announced its acquisition of Zivver in June 2025, so Zivver is now US-controlled. It appears in 73 entries at 73 organisations in the Dutch algorithm register, more than any other US-controlled supplier.
-
Yes, if the data is in Microsoft's possession, custody or control. The CLOUD Act lets US authorities order a provider under US jurisdiction to disclose such data wherever it is stored. Encryption with keys the provider cannot reach limits what it is able to hand over.
-
The national register of algorithms used by Dutch public bodies, published at algoritmes.overheid.nl. Registration is voluntary. Each entry describes purpose, data, risks and supplier. There is no field for the hosting provider or the AI model.
-
Yes. The CSV on this page has one row per register entry with Lindstead's classification and a link to the entry. Lindstead's analysis is published under CC BY 4.0. Cite Lindstead (2026) with a link to this page.
Sources
- Algoritmeregister (algoritmes.overheid.nl), full export via 'Download alle algoritmes', downloaded 4 October 2026: 1,574 entries from 341 organisations. algoritmes.overheid.nl/nl/algoritme
- TNO, Overheidsbrede Monitor Generatieve AI (TNO 2025 R12653), for the Ministry of the Interior, December 2025. open.overheid.nl/documenten/dafe19ab-8874-42a0-a55f-315ac5825282/file
- Computable, Overheidsautomatiseerder xxllnc pareert kritiek op AWS-migratie en wijst op Haven, 7 January 2026. www.computable.nl/2026/01/07/overheidsautomatiseerder-xxllnc-pareert-kritiek-op-aws-migratie-en-wijst-op-haven/
- Ministry of the Interior (BZK), Algoritmeregister publication standard, chapter 2, field 'Leverancier'. github.com/MinBZK/Algoritmeregister-Publicatiestandaard/blob/main/ch02.md
- Kiteworks, Kiteworks acquires Zivver, expanding private data network with zero-knowledge encryption, 17 June 2025. www.kiteworks.com/company/press-releases/kiteworks-acquires-zivver-expanding-private-data-network-with-zero-knowledge-encryption/
- Binnenlands Bestuur, Amerikaans bedrijf koopt aanbieder veilige communicatie, 23 June 2025. www.binnenlandsbestuur.nl/digitaal/amerikaans-bedrijf-koopt-aanbieder-veilige-communicatie
- Tweede Kamer, answers to questions by Kathmann on the takeover of Zivver by an American company, 16 December 2025. www.tweedekamer.nl/kamerstukken/kamervragen/detail?id=2025Z17137&did=2025D52592
- Reveal, ZyLAB is now Reveal, 2024. www.revealdata.com/blog/zylab-is-now-reveal
- Algemene Rekenkamer, Het Rijk in de cloud, 15 January 2025. www.rekenkamer.nl/documenten/2025/01/15/het-rijk-in-de-cloud
- Algemene Rekenkamer, Focus op AI bij de rijksoverheid, 16 October 2024. www.rekenkamer.nl/documenten/2024/10/16/focus-op-ai-bij-de-rijksoverheid
- Computable, Belastingdienst zet door met Microsoft 365, geen Europees alternatief beschikbaar, 2 December 2025. www.computable.nl/2025/12/02/belastingdienst-zet-door-met-microsoft365-geen-europees-alternatief-beschikbaar/
- 18 U.S.C. 2713, Required preservation and disclosure of communications and records (CLOUD Act). www.law.cornell.edu/uscode/text/18/2713
- Herziening rijksbreed cloudbeleid 2026, 3 July 2026, sections 1, 3.4, 4.5 and 4.6. www.tweedekamer.nl/downloads/document?id=2026D35295
- Motion Kathmann, Kamerstukken II 2024/25, 26 643, no. 1315, adopted 18 March 2025. zoek.officielebekendmakingen.nl/kst-26643-1315.html
- Ministry of the Interior (BZK), Kamerbrief over doorontwikkeling algoritmekader en register, 9 September 2025. www.rijksoverheid.nl/documenten/kamerstukken/2025/09/09/kamerbrief-over-doorontwikkeling-algoritmekader-en-register
- Autoriteit Persoonsgegevens, Rapportage AI- en Algoritmerisico's Nederland (RAN), edition 6, 2026. www.autoriteitpersoonsgegevens.nl/system/files?file=2026-03/ran-6-rapportage-ai-algoritmes-nederland.pdf
- Algorithm Audit, Inventory of high-risk AI systems, 25 August 2025. algorithmaudit.eu/knowledge-platform/knowledge-base/inventory_high_risk_ai_systems
Method: all register figures from Lindstead's analysis of the Algoritmeregister export of 4 October 2026; context sources opened on the same date. This page is general information, not legal advice. Corrections: contact@lindstead.com.