Language models and other AI services have long ceased to be an experimental playground in many companies. They answer customer enquiries, classify documents, generate text modules and support development teams in writing and reviewing code. As soon as such functions are anchored in business-critical processes, the risk situation changes: a service that was a gimmick yesterday is part of value creation today.

This is precisely where the current discussion comes in. As AI becomes productive, dependence on individual providers grows – and with it the question of how resilient one's own infrastructure actually is. Multi-sourcing, open-weight models and European offerings are cited as remedies.

Why provider lock-in with AI has a different effect

Dependence on suppliers is nothing new in IT. With AI services, however, several special characteristics come together that make switching harder than with classic software.

  • Behaviour instead of interface: Two models can serve the same technical interface and still respond differently. Prompts, i.e. the instruction texts to the model, are often optimised for a specific model.
  • Quality is hard to compare: Whether an answer is "good enough" cannot be measured as clearly as a database query. Without your own test cases, there is no basis for objectively evaluating alternatives.
  • Data and context: Embedded knowledge bases, vector indices and fine-tuning are often tied to the respective platform.
  • Pricing and model policy: Models are discontinued, replaced or restricted in their use. Anyone with only one source of supply bears these changes without mitigation.

Multi-sourcing: more than a second contract

Multi-sourcing means being able to obtain the same function in principle via several providers. The real value lies less in the second contract than in the architecture that makes switching possible in the first place.

In practice this means: the application does not talk directly to a provider's interface, but to its own intermediate layer. This layer encapsulates authentication, model selection, retries and logging. If a provider is replaced, ideally only the configuration changes – not the business code.

This includes a second, often underestimated building block: a collection of realistic test cases from your own use case. Only with these can it be assessed whether an alternative model meets the business requirements. Without this basis, every provider change remains a gut feeling.

Open-weight models as an additional option

Open-weight models are models whose trained parameters – the weights – are publicly available. In principle, they can therefore be operated on your own or rented infrastructure instead of being used exclusively as a third-party service.

The benefit lies above all in control: operation is independent of whether a provider changes its prices or switches off a model. Data does not necessarily leave your own environment. And the model version remains stable for as long as your own operation requires.

This is offset by real costs. Operation, scaling, updating and securing then fall within your own area of responsibility – including the necessary hardware and staff. For many mid-sized organisations, the relevant question is therefore less whether everything is operated in-house, but which sub-areas should be operated in-house at all. A mixed setup is often more viable: sensitive or particularly stability-critical tasks internally, broad or compute-intensive tasks via external services.

Keeping European offerings in view

European providers additionally broaden the choice. They are relevant not only for reasons of data processing within familiar legal jurisdictions, but simply as a further option in the supplier landscape. The more serious alternatives there are, the lower the pressure to accept a single platform without any fallback.

What matters here is a sober assessment: origin alone does not replace an examination of performance, availability and contractual terms. A provider belongs on the shortlist if it meets the business requirements – regional anchoring is an additional argument, not a substitute criterion.

What can be prepared concretely

For project managers, this results in a number of practical starting points that can be implemented regardless of the final provider decision:

  1. Stocktaking: Which processes already use AI services, and what happens if one of them fails for a week? This question separates convenience functions from genuine dependencies.
  2. Introduce an abstraction layer: Bundle provider calls in one place instead of distributing them across the entire application.
  3. Build test cases: Twenty to fifty typical inputs with expected results are often enough to make models comparable.
  4. Read contracts and terms of use: Notice periods, announcement periods for model changes and regulations on data use are part of the technical risk assessment.
  5. Keep data portable: Source documents, prompt collections and evaluation results belong in formats that remain usable outside the respective platform.

None of these points requires an immediate fundamental decision.

They do, however, reduce the effort if such a decision becomes necessary later.

Classification for web and software projects

For web platforms and specialised applications, this primarily means: AI functions should be designed from the outset as interchangeable components, not as a hard-wired part of the application.The effort for a clean abstraction layer and dedicated test cases is manageable during the ongoing project, but significantly higher afterwards.Those who do this preparatory work decide on providers and operating models later from a position of freedom of choice – and not under time pressure.

Sources