Anyone who builds AI functions into a product or an internal process also decides on a new dependency.In many cases, models are used whose inner workings cannot be seen by the organisation using them and which cannot be adapted to its own requirements.This is precisely where the Sovereign AI Foundation (SAIF) comes in, a new initiative under the umbrella of the Eclipse Foundation.It aims to highlight open-source alternatives.
Information is currently scarce: the direction is known, but not every detail of the implementation.Nevertheless, the initiative is relevant for project managers because it addresses a topic that is increasingly appearing in tenders and architecture decisions.
What sovereignty in AI projects is about
Sovereignty here does not just mean the server location.It covers at least three levels that are often mixed up in projects:
- Data sovereignty – who controls where inputs, documents and logs flow and how long they remain there.
- Verifiability – whether it is traceable how a model was built and on what basis it responds.
- Adaptability – whether a model can be adapted to domain-specific characteristics, its own terminology or regulatory requirements.
With pure API services, i.e. models that can only be reached via a provider's interface, above all the first level can be regulated contractually.Verifiability and adaptability, by contrast, remain limited because training data, weights and change cycles lie with the provider.
Why dependency becomes expensive in practice
Technical dependency usually only becomes visible after introduction.Typical patterns that can be observed in web and software projects:
- A model is discontinued or replaced by a new version.Prompts and downstream logic that were tuned to the old behaviour suddenly deliver different results.
- Response quality and response times change without the organisation's own application having been adapted.Error analysis is then only possible to a limited extent.
- Price and terms changes affect functions that are by now firmly anchored in everyday work.A short-term switch is then hardly realistic.
- Compliance questions arise retrospectively, for example when a business unit processes personal or confidential content.
None of these risks fundamentally argues against commercial models.But they do argue for choosing the dependency deliberately and limiting it architecturally.
What open models can do – and what they cannot
Open-source alternatives shift the balance but do not automatically solve every problem.Open models can be operated on your own infrastructure or with a service provider of your choice, which keeps data flows manageable.They can be versioned and frozen so that an application's behaviour remains stable over a longer period.And they can be adapted when a company's specialist language, document structures or validation rules deviate from the general case.
In return, responsibility moves in-house: operation, updates, security and quality measurement must be organised.The term "open source" is also not used consistently in the AI field – some models publish weights but not the training data or the complete licence terms for commercial use.An initiative that provides orientation here therefore addresses a real gap.
Consequences for the architecture
Regardless of how the Eclipse initiative develops further, interchangeability can already be anchored in projects today.Four points are particularly helpful:
- Introduce an abstraction layer. Applications do not talk directly to a provider but to an internal interface.A model change then affects one component, not half the codebase.
- Build your own test cases. A manageable collection of real tasks with expected results makes it measurable whether an alternative model is sufficient for the specific purpose.
- Keep data and knowledge separate. Content, search indexes and rules belong under your own control, not in the fine-tuning of a third-party model.
- Assess criticality. A text suggestion in the editorial system has different requirements from information with legal effect.Not every use case needs the same sovereignty.
Classification for TYPO3 and software projects
In content management projects, AI functions now appear in many places: text suggestions in the editorial team, translations, alternative texts for images, semantic search, classification of enquiries.These functions are individually small but collectively business-critical – and they quickly end up with a single provider because that is the fastest route at the start.
For planning, this means: model selection should be a decision with an exit option, not a one-way street.A hybrid approach is pragmatic in many cases – commercial models where quality and speed matter, open models where data is sensitive or results must remain reproducible over years.
For web and software projects, the focus thus shifts from the question "which model is the best" to the question "how expensive is a switch".Anyone who plans abstraction, their own quality measurement and clear data flows from the outset can observe new developments such as the Sovereign AI Foundation without having to question the architecture every time.
And anyone who has to prove data sovereignty anyway has one more option on the table with verifiable and adaptable models.
Sources
- Less dependence on AI providers: Eclipse launches Sovereign AI Foundation — heise developer News