Artificial Intelligence (AI) is entering a new stage in which Europe could play a more significant role. On the back of large general-purpose models (LLMs such as ChatGPT, Claude and Gemini), the focus is shifting towards smaller, customized analysis models (SLMs) to meet the specific needs of businesses and regulated sectors. This is one of the key findings of the strategic monitoring report by Santander AI Lab, in cooperation with Fundación General CSIC, entitled “The New European Frontier of B2B AI”.
Key Trends
1. The era of “One Model Fits All” is beginning to run out
Large generalist models will continue to be relevant, but they are not the optimal answer for all regulated business processes. For recurring, document-based or highly regulated tasks, Small Language Models (SLMs) and specialised models offer a more efficient alternative. They make it possible to reduce inference costs, improve latency and strengthen the competitive advantage through greater control over data, weights and traceability.
2. Sovereign RAG becomes a key architecture
The separation between the model and corporate memory will be a structural decision. A regulated organisation does not need to retrain models using all its internal knowledge, but rather to connect controlled models with governed document repositories. This architecture makes it possible to use generative AI without losing control over case files, internal policies, regulatory documentation or customer data.
3. Regulation shifts from being a brake to a competitive advantage
EU AI Act, DORA, GDPR and NIS2 raise operational requirements, while also professionalizing the market. Entities capable of demonstrating traceability, human oversight, risk control, data quality and technological resilience will be better placed to scale AI in critical processes. In banking, rigorous compliance can act as a mark of trust.
4. Talent is no longer a support function; it becomes core infrastructure
Competitive advantage will no longer depend solely on hiring specialised technical profiles. The report points to a hybridisation between business experts, legal profiles, risk teams, technologists and European scientific networks. AI-driven differentiation in banking will be built on domain and business knowledge, not solely on model engineering.
5. Maturity and control must guide prioritisation
The report identifies RegTech, fraud prevention, AML, KYB, document AI, synthetic data, model governance and supervised agents as priority processes. These are areas where the Bank can capture efficiency, reduce operational risk, and build scalable capabilities without assuming excessive dependence on proprietary vendors.
The study delves into the current stage of development of AI in Europe and its potential impact on critical sectors. While large language models (LLMs) and consumer applications were at the forefront of the first wave of AI adoption, the strategic debate is now beginning to home in on more specialized, secure and governable uses. The study reveals what appears to be certain restraints of large general-purpose models, which are not always the best solution for all business processes. Conversely, small language models (SLMs) and specialized models stand out as a more efficient alternative as they reduce inference costs, improve latency and strengthen competitive advantage through greater control over data, weights and traceability.
According to José Manuel de la Chica, Head de Santander AI Lab, “it’s time to move on from experimenting with general-purpose models to industrial, sovereign and governable B2B AI. The answer won’t always be using the most powerful model, but rather in embedding the right AI into the right process, with traceability, data control and accountability from design”.
Europe has the opportunity to make a stand in AI: to compete in vertical specialization, regulatory trust, data sovereignty, and scientific quality. That approach is well suited to the banking sector because it lies at the heart of the business. Effective sovereignty does not mean isolation but knowing which processes must remain under control.
This paradigm shift is particularly significant for such a critical and regulated sector as banking, where technology rollout must follow strict requirements regarding traceability, resilience and privacy. That's why financial AI “is not being developed solely around chatbots and basic automation, but rather in risk management, fraud, creditworthiness, insurance, mathematical modelling, AI governance and responsible AI”.
Europe’s strengths
Against the backdrop of a global race dominated by major investments in foundational models (the US mobilized $109 billion in private investment in AI in 2024, compared with $19 billion in Europe), the report sets out a trailblazing path for Europe: to compete in vertical specialization, data sovereignty, operational efficiency, scientific quality, and regulatory trust. For a bank, this means “combining technological ambition with governance discipline, sound economic judgement, and an architecture designed to scale AI securely”.
Another of Europe’s strengths could be regulation itself, which would go from being a hindrance to a competitive advantage. So while the regulatory framework that includes the AI Act, DORA, the GDPR, NIS2 and others raises operational requirements, it also makes the market more professional by setting quality standards and creating a competitive barrier against technologies that are less transparent or more difficult to audit.
The study also analyses the role of Europe’s major innovation hubs (Paris, DACH, London and Spain) and the importance of infrastructure such as supercomputing, AI factories and data labs in building more secure, more efficient and better-controlled AI.
The report’s findings stem from the analysis of 10.3 million documents, with a further 326,000 AI, finance and banking publications poured over. The separation between the model and corporate memory is another trend that the report examines. In particular, the use of sovereign recovery-augmented generation (RAG) will make it possible to separate the AI engine from the corporate memory, thus preventing internal knowledge from being absorbed by external suppliers.
