Moving data does not eliminate silos. Installing an artificial intelligence (AI) assistant does not make documents more reliable. And having a high-performing supplier does not guarantee that you will be able to switch away from them. The authors advocate different commercial offerings or research approaches. But their analyses shed light on the same challenge: that of transforming technical capabilities into services that can be used over the long term. The issue concerns internal responsibilities just as much as the choice of infrastructure.
The first hurdle often lies before AI. The information exists, but remains scattered across applications, departments and formats. “The first question is no longer simply how much data a company has, but what it is capable of doing with it,” explains Malick Thiam, sales architect at Proximus NXT Luxembourg. His colleague Jad Houdeib, head of cloud consulting, describes the limitations of a migration carried out without giving due consideration to the data being transferred: “In practice, we still too often see cloud migrations carried out on an application-by-application basis, without giving due consideration to the data. This simply moves the silos rather than eliminating them.”
Business accountability at the heart of everything
Stéphane Zema, director and cloud transformation leader, and
Loïc Dunand, managing director and lead data & AI at PwC Luxembourg, present a complementary analysis. In their joint contribution, they place business responsibility at the heart of the framework. A platform can organise data exchanges and enforce access rules. It does not decide, on behalf of the organisation, who is responsible for ensuring the quality of information or approving its use. They therefore recommend proceeding on a business-area-by-business-area basis, starting with use cases where the value can be clearly identified.
This organisation takes concrete form at local authority level. The Inter-municipal IT Management Association (Sigi) explains that it is cataloguing the data from its applications and describing it using a shared vocabulary, drawing on the national LetzData catalogue. The aim is for the same information to be understood in the same way by several public authorities. As part of this initiative, data exchange interfaces established with Esch-sur-Alzette and Dudelange have been operational since November 2025. Their implementation requires specifying the information that is accessible and the conditions under which it is transmitted.
“Sigi facilitates these exchanges whilst ensuring that local authorities retain control over their data and the conditions under which it is shared,” says Johny Nevoa, acting director of the union. For residents, the expected benefit lies in a simplified process or in no longer having to enter the same information several times. Here, data processing serves a purpose that is independent of any promises made by AI.
Being confident in your rules and processes
The advent of AI assistants is making this field more demanding. An out-of-date document can form the basis of a convincing response. An authorisation that is too broad can make information accessible that should have remained restricted. Automation therefore exacerbates the consequences of a pre-existing flaw.
At NSI Luxembourg, Olivier Baltus, a technical leader IBM, describes this issue using AskNSI, an internal assistant accessible via Microsoft Teams. According to his post, the assistant helps staff find internal procedures and rules. For it to work properly, it is necessary to check that the documents used are still up to date. When a policy changes, the business unit must approve the new version, after which the technical teams must ensure the assistant takes it into account.
User feedback is not enough either. An error must be analysed to determine whether it stems from the document itself, the information search process or technical issues. This analysis must then inform the tests designed to prevent it from recurring. “It is this ongoing work that transforms a prototype into a reliable service that staff can rely on in their day-to-day work,” emphasises Olivier Baltus.
Access control links this experience to cybersecurity concerns. Sébastien Bourgasser, dartalis’s CTO, draws attention to agents capable of viewing documents or emails and carrying out tasks. Their usefulness depends on the permissions granted. So does the risk they pose.
Franck Wurtz, the director of services at dartalis, puts it this way when discussing the limitations of a purely declaratory approach: “An internal policy is not enough if the organisation cannot control access and data exchanges.”
Hierarchies based on earnings
For managers, these requirements come at a cost that goes beyond simply purchasing a licence or using computing resources. They need to maintain the data sources, monitor responses, manage incidents and keep the connections with applications up to date. Robby Melin, sales director at NSI Luxembourg, therefore recommends prioritising projects based on their expected benefits and feasibility. “For senior management, the AI strategy must not become a series of isolated experiments. It must form part of a coherent roadmap, with clear priorities, measurable results and appropriate governance.”
Ludovic Gilles, Deep’s chief commercial officer, adds an economic caveat to this organisational requirement. “AI, today, must demonstrate its value in the real economy.” He observes that the resources required for industrialisation can dampen the enthusiasm generated by a pilot scheme. Simply completing a task more quickly is, therefore, not enough to establish the service’s profitability once it is rolled out. The benefits must be weighed against operating costs and the necessary organisational changes within the company.
This economic trade-off is linked to that of sovereignty. The contributions do not advocate a single environment for all uses. Hector Dumas, business consultant, and
Nicolas Mansuy, a senior business consultant at BearingPoint, suggests classifying applications and services according to their criticality, the sensitivity of the data and the consequences of a disruption. This approach leads to tighter controls where a dependency would pose the greatest risk. “The aim is not to ensure sovereignty across the board, but where dependence is greatest,” summarises Nicolas Mansuy.
Between sovereignty and compliance with the rules
The location of the hosting facility is therefore just one criterion amongst others. The contributors also recommend examining the operating conditions, subcontractors, the technologies used and the options for replacement. PwC emphasises in particular that a migration may require moving applications and their processing logic, rather than just the data itself. BearingPoint stresses the need for in-house expertise to make this transition feasible.
In the financial sector, part of this requirement is laid down in law. Article 28 of the European Regulation on Digital Operational Resilience (Dora) requires the entities concerned to have contingency plans in place for IT services supporting critical or important functions. These plans must be documented, adequately tested and reviewed periodically. The Regulation also requires the identification of alternative solutions and the preparation for the transfer or reintegration of services.
Loïc Didelot, CEO of Datacenter.eu Luxembourg, suggests a step-by-step approach to this concept: an independent backup, an additional connection or a disaster recovery environment can complement the existing setup. “The right time to look for a second provider is when the first one is working perfectly.” This diversification addresses the risk of concentration. However, BearingPoint’s analysis specifies one condition: one must also examine the shared dependencies behind the direct service providers.
Physical infrastructure remains essential to these decisions. Paul Konsbruck, CEO of LuxConnect, points out that their development involves connectivity, electrical capacity, cooling and expertise. “Digital sovereignty is not built around a single technology or project. It results from the establishment of the appropriate infrastructure, resources and skills, as well as their continuous development in line with changing needs.”
The “à la carte” aspect to be assessed
The AdwäisEO experience illustrates another aspect of this investment. Giovanni Corato, technical director, and
Lucien Hoffmann, executive director, describe an infrastructure combining various storage technologies for Earth observation data. According to their paper, the choices vary depending on computing, archiving and backup requirements. The company also states that it has a backup site in France. Data management can, therefore, be accompanied by geographical distribution designed to enhance business continuity.
The allocation also applies to salaries. Djamel Khadraoui, head of the Reliable Distributed Systems unit at the Luxembourg Institute of Science and Technology (List), argues that the cloud and computing carried out close to the devices – known as “edge computing” – are complementary. The former enables large-scale analysis. The latter can support a local response and the continuation of certain operations when connectivity is limited. However, this distributed approach brings its own constraints: more devices to protect, more models to update and more rules to enforce.
“In distributed AI, the risk is also distributed,” the researcher points out. His observation builds on the central issue of this report: every additional capability creates operational responsibilities. For management, therefore, the choice of a platform involves far more than just an IT budget. It involves the organisation’s ability to correct information, restrict access, maintain a service and switch providers when necessary.



