Depending on a handful of AI vendors: where things stand in 2026 and what you can do
· By Bee Flow · 8 min read
Most of the AI that Dutch companies use comes from a few American firms. What the past year showed about that risk, what The Hague and Brussels say about it, and the steps a small or mid-sized company can take itself.

On 6 November 2026 the Dutch cabinet adopted its vision on AI, titled Kiezen voor AI voor iedereen ("choosing AI for everyone"). The government's announcement contains a line any business owner could apply to their own company: "We do not need to be able to do everything ourselves, but we do want to decide for ourselves which way we go." The cabinet wants to keep the Netherlands from becoming too dependent on a small number of companies for important technology. The same announcement says it wants to remove the obstacles that keep small and mid-sized businesses from using AI.
Those two aims pull against each other a little. The more AI you use, the more you depend on your vendor to keep delivering: on the same terms, with the same model, without an outage. This piece is about that dependence. How large it is, what happened over the past year, what The Hague and Brussels say about it, and what a small or mid-sized company can do about it itself.
How concentrated the market is
Start with the cloud, because nearly all AI runs on it. According to Synergy Research Group (July 2025), Amazon, Microsoft and Google together hold 70 percent of the European cloud market. European providers fell from 29 percent in 2017 to 15 percent in 2022 and have stayed at roughly that level since.
Government is no exception. In January 2025 the Netherlands Court of Audit (Algemene Rekenkamer) reported that more than half of the central government's important public cloud services were bought from those three American companies. For 84 of the 126 services it looked at, the required risk assessment was missing.
For language models we found no market figures of the same quality, so we leave them out. What is certain is that ChatGPT, Copilot and Gemini, the names heard most often in an office, come from American companies.
Use is growing fast. Statistics Netherlands (CBS) reported in December 2025 that 27 percent of Dutch companies with 10 to 50 people use AI, up from 11 percent in 2023. For 50 to 250 people it went from 20 to 45 percent over the same two years, for the smallest firms from 7 to 14 percent. A company that did nothing with AI two years ago may by now have built a customer process on it.
What the past year showed
You only notice a dependence when something changes. Over the past year, four kinds of change came by.

Models have an end date. Anthropic keeps a page listing which Claude models are being retired, and promises customers at least 60 days' notice. On 30 September 2026, users of Claude Sonnet 4.5 were told the model leaves the API on 30 November 2026. Earlier this year Claude Sonnet 4 and Opus 4 went: announced on 14 April, gone on 15 June. After that date every request to the old model fails. This is how the market works, at every vendor: new models arrive and old ones leave. For you it means that a prompt, an agent or an automation tuned to one specific model has to be tested again by a fixed date.
Outages hit many customers at once. On 18 November 2025 at 11:20 UTC, Cloudflare's network started failing to deliver core traffic, and only at 17:06 was everything working again. In its own write-up Cloudflare called it its worst outage since 2019, caused by a changed database setting and not by an attack. A month earlier, on 19 and 20 October 2025, AWS's Northern Virginia region was affected for more than fourteen hours by a fault in DynamoDB's DNS management, according to AWS's own summary. If all your AI work runs through one chain, all of it stops when that chain does.
Prices and bundles change. On 4 December 2025 Microsoft announced new list prices for its commercial Microsoft 365 suites from 1 July 2026, together with more Copilot Chat in those suites, such as awareness of your inbox and calendar. An assistant built into your office software comes in the shape the vendor picks, at the moment the vendor picks. That is convenient while it fits. It also makes switching harder, because mail, documents and assistant become one bundle. How Bee Flow compares with Copilot is on Bee Flow vs Microsoft Copilot.
The owner of your vendor can change. The clearest Dutch example of the past year is Solvinity, the IT company DigiD (the Dutch government's login system) depends on. On 27 January 2026 the Dutch House of Representatives held a round-table session on what it would mean for DigiD if the American company Kyndryl bought Solvinity. On 25 May 2026 State Secretary Aerdts prohibited the takeover, on the advice of the Dutch investment screening office (BTI), which concluded that it "possibly poses a risk to the public interest", according to the letter to parliament of 26 May. The same letter stresses that the screening is country-neutral. Solvinity has objected (letter of 17 June). A small business has no investment screening behind it, but the question is the same: what happens to your data and your contract if your vendor has a new owner tomorrow.
Vendors plan for the worst case too. On 30 April 2025 Microsoft wrote that if it is ever ordered to suspend its European cloud operations it "will promptly and vigorously contest such a measure using all legal avenues available", and that it keeps back-up copies of its code in Switzerland. When a vendor writes that scenario into its contracts, that is reason enough for a customer to think about it as well.
What The Hague and Brussels say
The Court of Audit was clear about the risk in January 2025: there is a risk that foreign governments, particularly the US government, can access and even modify information about the Dutch government and its citizens. Its main recommendation was that central government should face the large cloud providers as a single organisation.
In its new AI vision the cabinet does not aim for complete independence. The goal is enough knowledge and capacity of its own to keep access to important technology, and to choose its own direction.
In Brussels, the European Commission presented a tech sovereignty package on 3 June 2026. In it the Commission writes that "Europe remains heavily dependent on suppliers outside the European Union for core digital technologies". The main legislative proposal is the Cloud and AI Development Act. According to the European Parliament's summary it sets up a four-level Union assurance scheme for cloud services. Public authorities working on public order would only be able to buy services at level 2, 3 or 4, and each member state would designate at least one data centre acceleration zone. On 25 October 2026 the proposal was still waiting in Parliament for a decision on the committee that will handle it. So it is a proposal, not law. We would wait for the final text before building anything on it.
Industry is pushing too. In March 2025 European companies writing under the EuroStack banner sent an open letter to Commission President von der Leyen. They asked for "a formal requirement for the public sector to 'Buy European'", adding that "the aim is not to exclude non-European players".
The large American providers answer with European versions of their own. On 14 January 2026 AWS opened the AWS European Sovereign Cloud in Brandenburg, Germany: physically and logically separate from the rest of AWS, managed through European legal entities under German law, and according to AWS to be operated exclusively by EU residents located in the EU. A Local Zone in the Netherlands is planned. Whether such a setup answers the Court of Audit's concern about foreign access is a legal question, and the answer depends on the situation. This article is not legal advice.
What a small or mid-sized company can do

You will not solve this on your own, and you do not have to. A realistic goal is being able to switch within a few weeks when you must, rather than within a few years. That starts with a few habits.
- Make a list of the AI services your company uses, which data goes into them, whose name the account is in and which work depends on them. There are usually more accounts than anyone thought.
- Pick a model per task. Writing a title, sorting a mail or summarising a short message works fine with a small model. A hard analysis needs a large one. Split tasks this way and the retirement or outage of one model only touches part of your work.
- Look at European hosting and open-weight models. Mistral, Scaleway and IONOS serve models from European data centres, see EU-hosted LLM providers. An open-weight model is offered by several hosts and can also run on your own hardware. Then you can change host without changing model.
- Keep what is yours with you. Store your instructions and standard prompts in your own documents. Leave your knowledge sources in your own storage instead of only uploading them into one vendor's assistant. And build automations in a form you can export. In Bee Flow an automation exports as a JSON file and imports into another installation as a new draft.
- Test a replacement before you have to. Write down a small set of standard questions with answers you are happy with, and have a second model answer them now and then. When a retirement notice arrives, the test is ready. In Bee Flow you can mark an automation run you trust as a golden run and replay it after a change such as a new model, to see where only the AI output differs (the platform).
- Ask your vendor about the exit: can you export your data, how much notice do you get of changes, and what happens on a takeover? More questions are in five questions for the vendor.
Where Bee Flow fits
We built Bee Flow so that you can change the model without changing how your team works. The same workspace talks to Claude, OpenAI, Google, Azure, Mistral, Scaleway, IONOS and EU GPT, and to models on your own hardware through Ollama, vLLM, llama.cpp and other runtimes. An administrator picks a model per tier, a small one for titles and sorting and a large one for hard analysis, and an agent can have its own model. When a model is retired or a provider is down, you change that setting, and the agents and automations that use the tier pick up the new model. A knowledge base can read folders from Nextcloud, OneDrive or Google Drive, so your sources stay in your own storage. You can host Bee Flow yourself or have us run it in the EU, and the free licence runs on your own server without a key (what running AI on your own server takes). The list from step 1 is a good place to start, and if you want to go through it with Bee Flow in mind, get in touch.