home / knowledge base / ai without lock-in
knowledge base

AI Without Lock-In: What DNB Tells SMEs_

The Dutch central bank (DNB) advises companies not to depend on one AI provider that is hard to replace. That sounds simple until you ask what has to stay swappable. The answer is less the model than your data.

what_dnb_said

What DNB said on 2 November

On 2 November 2026 De Nederlandsche Bank (DNB) published an analysis of the economic impact of AI in the Netherlands. Three points matter for SMEs. Smaller companies use AI less often than large ones, as do companies in construction, transport and hospitality. Smaller companies in particular can benefit from software and cloud services that make specialist knowledge accessible, without having to build their own AI models. And according to DNB, AI does not lead to higher productivity by itself: you have to choose applications that fit, train employees and adapt work processes.

DNB also says that companies need reliable access and should avoid depending on one provider that is hard to replace. That is a general recommendation, not an SME manual. The text contains no SME figures, so we do not invent any. What is useful is that the recommendation points to a question that is rarely asked in practice.

what_is_hard_to_replace

What is actually hard to replace?

An AI model is easier to replace than you think. Models improve and change price, and many applications can run on another model with a different setting. It gets harder with everything around it. Think of it as three layers, where the bottom one weighs the most.

what can you swap, and what does it cost
layer_03

The AI model

Usually swappable with some configuration and a test round. This is where the least lock-in sits.

layer_02

The application and your way of working

The assistant, the integrations, the agreements about who checks what. Swapping is possible, but costs work and getting used to it.

layer_01

Your business data

Customers, quotes, projects, the history. If it is stuck in one package, everything above it is stuck too.

three layers under AI: the lower the layer, the more expensive the swap
where_it_gets_stuck

Lock-in rarely comes through the model

Lock-in usually arises because the AI feature sits in the package where your data already lives. That is convenient, because the assistant sees your customer data directly. It also means the AI only works on what is in that package, and that your data, including everything the AI adds to it, has to come along when you leave. That gives you a new reason to stay, on top of the contractual ones.

Anyone who has missed a renewal date knows the contract matters. We wrote about that in the renewal clause in your CRM contract. With AI there is an extra layer: not only the data you entered, but also the summaries, labels and scores the system created itself.

three_questions

Three questions to ask a vendor

You do not have to be a lawyer or an architect to test this. Ask these questions before you switch on an AI feature in your package:

  • Can I get everything out, including what the AI added? Ask for an export in a common format, including summaries, labels and scores.
  • Does the AI also work on data outside this package? If not, your overview is only as wide as this one system.
  • What happens to my settings if the model or the vendor changes? Prompts, rules and approval steps are work you do not want to do twice.

If the answer to all three is yes, you are in good shape. If one of them is no, at least you know what you are buying.

when_it_does_not_matter

When you need not worry about this

DNB also advises starting with an application that fits and a process you can adapt. For a first trial, an AI feature in the package you already have is often the sensible choice, even if your data stays there. An in-house data layer is not a precondition for starting with AI. It becomes interesting once you want several applications to work on the same customer data, or when you notice you cannot leave without a big move.

So this is not about reorganising everything now. It is about knowing where you are locked in, before it becomes a problem.

and_now

What to do tomorrow

Write down which AI features you use now and in which package the related data lives. Then run through the three questions for the package that holds most of your business data. It takes an hour and gives a more honest picture than any vendor brochure.

If you do not yet know where to start with AI, read how to spot repetitive work. Why your data is the foundation is explained in what is a data layer, an honest answer, and why many companies use AI mainly for text in SMEs use AI for text, but rarely for their own numbers. How we approach this, with your data central in an open data layer and software you can swap, is described on the AI Native Data Layer page.

frequently_asked_questions

Frequently asked questions about AI and dependency

What does DNB mean by dependence on one AI provider?

DNB advises companies to secure reliable access to AI and to avoid depending on one provider that is hard to replace. For an SME, the practical point is to make sure your data and your way of working can move with you when you change model or vendor.

Do I need my own data layer to start with AI?

No. For a first process, an AI feature in your existing package can be enough. A data layer becomes relevant when several applications need the same customer data or when you can no longer switch without a big move.

Keep reading: the whitepaper

The whitepaper on the AI Native Data Layer explains how your data stays yours while AI models and software change, and how to start small.

Download the whitepaper