The desire for digital sovereignty is not just being driven by compliance departments. More and more companies see dependence on non-European providers as a risk. According to the GenAI Impact Report Germany 2025, 71 per cent of respondents now consider it important or very important that their company’s GenAI applications were developed in the EU. In terms of data location, as many as 85 per cent consider a data centre on European soil, protected from possible interference by foreign governments to be crucial. In addition to data protection and regulatory requirements under the EU AI Act, the focus is on controlling the flow of information, model transparency and the long-term cost structure. Companies should consider the possibility of making the switch, even if not every AI has to be “Made in Europe” in the future. For most companies and for many application scenarios, the use of language models and corresponding services from global players is entirely unproblematic. With a view to prevention, however, the key question is: How can a switch be made if it becomes necessary?
From adesso’s point of view, there are five key areas of action:
- Starting point: Evaluate the situation with an AI map. Before switching to an alternative GenAI solution can even be assessed, a complete overview is required: Which GenAI services are in use? Which interfaces do they use? Which models are in the background? How is the data stored? This AI map creates transparency about existing integrations in systems, tools and other data pipelines. It also reveals where shadow IT is involved. It is advisable for an interdisciplinary team from IT, data protection, specialist departments and compliance to jointly classify data according to criticality, data relevance and strategic importance.
- Evaluation: Compare language models. Not every model is suitable for every application. Companies should therefore define in advance which requirements apply in each case – for example with regard to language, accuracy, hosting, licence conditions or regulatory compliance. A test environment should then be set up. Ideally, the comparison is carried out under real-world conditions: same prompts, same context and defined evaluation criteria such as consistency of results, hallucination rate or response time. The technical evaluation is supplemented by a licence assessment and an analysis of governance. In this step, a targeted comparison of open source models and proprietary providers should be carried out, as this decision has an impact on the entire solution architecture.
- Migration planning: Set up a roadmap with a step-by-step model. A successful exit from existing LLM dependencies cannot be achieved with a “big bang”. A step-by-step model that starts with non-critical use cases is a proven approach. For example, the change can start with internal text-based tasks or knowledge systems – such as an RAG (Retrieval Augmented Generation) system based on a sovereign model. At the same time, existing vector databases that give language models access to relevant company data can be abstracted so that the backend can be replaced without having to adapt the user logic. It is important to plan fallback options in each phase, for example through parallel operation. It also makes sense to have an internal centre of excellence in which expertise is pooled centrally and which supports the rollout.
- Infrastructure: Decision on operating model is central. Whether or not language models run efficiently and securely depends largely on the infrastructure. Many European alternatives rely on on-premises or private cloud scenarios, which has advantages in terms of data protection, but places new demands on hardware, deployment and monitoring. Companies must therefore decide at an early stage whether to operate their own GPU clusters, use European IaaS providers or work with specialised service providers. It is important to note that the infrastructure must not only be scalable, but also compatible with existing workflows, APIs and security standards. Otherwise, every exit strategy risks remaining in pilot status.
- Acceptance and expertise: People remain the key factor to success. Every technological migration needs the support of the organization. Expectations are high, particularly in the area of GenAI. At the same time, trust is fragile. Companies should therefore involve internal pilot users at an early stage to test the models for relevance and comprehensibility. Training formats, guidelines for prompt generation and clear feedback channels ensure acceptance and increase the level of maturity of use. It is also important to clearly communicate internally why the migration is taking place – not as an end in itself, but to strengthen the company’s ability to innovate and its digital resilience.
