In more than nine out of ten cases, the use of artificial intelligence is for internal purposes. Photo: Shutterstock

In more than nine out of ten cases, the use of artificial intelligence is for internal purposes. Photo: Shutterstock

A joint report by the Luxembourg Central Bank (BCL) and the Financial Sector Supervisory Commission (CSSF) highlights a growing adoption of AI in the country's banks and financial institutions. From the boom in generative AI solutions to employee training initiatives and the challenges of governance and regulatory compliance, an overview of a revolution in the making.

The joint survey conducted in summer 2024 by the BCL and the CSSF--which received responses from 461 institutions, representing a participation rate of 86%--shows that around 43% of the institutions surveyed are now using AI solutions under development or already deployed in production, compared with just 30% in the previous survey. In total, 36% of respondents reported at least one concrete use case for AI within their organisation, representing 402 identified use cases, more than half of which are already operational in production. Significantly, nearly all (92%) of these use cases are for internal use (process optimisation, employee tools, etc.) rather than for direct customer services--a sign that, for the time being, AI is primarily being used to improve efficiency behind the scenes.

This adoption of AI is being driven by the efficiency and productivity gains that financial players expect from it. The top three benefits highlighted by respondents are improved internal processes, optimised operations (and reduced costs) and the ability to analyse vast volumes of data. In mirror image, the major obstacles to AI implementation relate to data: data quality is cited as the number one challenge, followed by data protection and data governance.

In terms of investments, the survey indicates that in 2024, almost half of entities invested in AI (or distributed ledger technologies - DLT) via their parent company or group (46% of respondents), while few conducted AI investments specifically at a local level (4% local only, and 9% combining local and group efforts). Even more significantly, over a third of institutions (36%) had not committed to any investment in these innovative technologies in 2024 - mainly among fund managers and some investment companies, while banks and payment institutions were more active. However, this lag in local investment may only be temporary: for the period 2025-2026, a clear acceleration is anticipated, with a predicted increase in AI investment well ahead of that in blockchain, particularly at the level of local Luxembourg entities (with the generative AI segment predicted to be the main driver of this growth). At the level of international financial groups, AI spending is expected to remain stable, which can be explained by the fact that many developments have already taken place at parent company level; the challenge now is to spread these advances to local entities, capitalising on the experience gained at group level.

A significant proportion of the players surveyed are experimenting with AI but have not yet deployed it on a large scale, which suggests that current pilot projects will be ramped up to production solutions in the near future. The recent rise of certain consumer technologies is also accelerating this phenomenon, by democratising access to AI tools without the need for heavy in-house development.

The rise of generative AI is transforming use cases

Since the availability of advanced conversational and creative models at the end of 2022 (in the image of ChatGPT, to name only the best-known), the adoption of AI has experienced a global acceleration, and the Financial Place is no exception to the rule. The BCL-CSSF report even points out that the proportion of financial entities that have developed generative AI use cases now exceeds that relying on traditional machine learning or other forms of AI. More concretely, 61% of all AI use cases identified in the survey incorporate generative AI solutions - whether text-generating language models, image creation tools, or other applications - confirming the sector's appetite for these new technologies capable of autonomously generating content.

New use cases have emerged or are gaining momentum, such as automated document synthesis (summaries of reports, financial news), content generation (e.g. assisted writing of reports or analyses), next-generation conversational chatbots for customer service, AI-enhanced machine translation, and code generation to aid software development. These categories, once experimental, are now among the most widely implemented by institutions in Luxembourg's financial centre.

However, "traditional" AI is not left behind. Machine learning techniques continue to play a central role, particularly in risk management and compliance functions. Use cases such as the detection of fraudulent transactions, the fight against money laundering (AML) or the monitoring of suspicious transactions in relation to the financing of terrorism are largely equipped with proven ML algorithms. Often, moreover, these algorithms are now enhanced by advances in generative AI - think of surveillance tools that combine traditional predictive models and natural language analysis to detect weak signals in documentation or correspondence. More innovative uses, meanwhile, are beginning to emerge in business functions: for example, in-house virtual assistants capable of helping employees by synthesising information or coding short scripts.

Credit scoring (automated assessment of creditworthiness), which is one of the few use cases explicitly classified as "high risk" by future European regulations, remains relatively limited in Luxembourg for the time being.

Training, governing and securing: AI increasingly framed

According to the report, the vast majority (84%) of survey participants have already implemented or plan to deploy AI training programmes for their employees - ranging from basic awareness modules on AI concepts to advanced technical training for specialists.

In parallel, almost half of institutions say they have formalised their approaches by adopting internal AI policies: 43% of respondents have an approved AI policy (whether a general policy incorporating principles on AI, or a policy dedicated specifically to AI). While still modest in absolute terms, this figure nevertheless represents a marked increase on the previous survey, reflecting the growing attention being paid to the ethical and organisational issues surrounding AI. More institutions are adopting guidelines on the use of AI, the management of bias, or even the validation of models before they go live. In fact, compared to last year, more institutions claim to have implemented ethical policies or algorithmic bias detection/prevention techniques, a sign that they are increasingly integrating the reliability and fairness dimensions of AI into their projects.

More than one financial company in two (54%) has already implemented specific IT security measures to counter AI-specific vulnerabilities (e.g. protection against adversarial attacks on models, drift monitoring). This is a significant jump compared to the previous year, when this proportion was much lower - an increase that even reaches 66% if we only consider the scope of banks, payment and electronic money institutions, i.e. the players already covered in the first survey.

An increase in skills

Nearly two-thirds (63%) of entities using AI have a dedicated data science team to develop and monitor these projects. More often than not, these specialist teams are pooled at group level (in 55% of cases) rather than established locally, reflecting a strategy where technical expertise is centralised at head office and its fruits disseminated to subsidiaries. A few institutions (3%) have nonetheless set up AI teams directly in Luxembourg, while a significant proportion have no in-house team at all - mainly because turnkey solutions such as GenAI make it possible to experiment without mobilising specialised in-house skills. Where they do exist, the data science teams remain modest in size (often fewer than 10 people) and are mainly attached to IT departments, although around a third report directly to the business or in a mixed IT-business way. This governance of AI by technology, with an increasingly close link to business needs, illustrates the desire to integrate AI into the organisation in a pragmatic and controlled way.

Finally, the human element remains more than ever in the loop. The report's statistics show that the systems in place provide almost systematically for human intervention or validation at some stage of the AI-powered process, a token of caution in a context where ultimate responsibility must remain under control. This "augmented AI" model is also in line with regulators' recommendations for trustworthy AI.

This article was originally published in French.