From experimentation to useful, controlled and industrialised AI.

From AI prototype to operational service

Companies are no longer short of ideas when it comes to artificial intelligence. The real challenge now is to turn these ideas into pilot projects, and then into services capable of creating long-term value. This involves linking AI to business processes, data and enterprise applications, whilst incorporating security, governance and operational requirements.

A successful demonstration does not guarantee that a solution will work in day-to-day use. In a production environment, an AI service must provide relevant answers, adapt to changes in the data, keep costs under control and allow for intervention when it makes a mistake or stops responding. AI cannot therefore be treated as a stand-alone tool: it requires businesses to rethink their operations across the entire organisation.

At NSI Luxembourg, our aim is to support this journey from idea to production, starting with a key question: where can AI genuinely improve a process, the employee experience or customer service, and under what conditions?

Prioritise projects that create value

Before choosing a model or platform, it is important to start by identifying the organisation’s actual needs. We help organisations identify and prioritise their projects by assessing expected value, feasibility, data availability, integration with existing processes and security constraints.

We prioritise well-founded ‘business cases’ over use cases chosen simply because they are ‘trendy’. This involves analysing the costs of current processes, identifying bottlenecks and estimating potential gains before launching a pilot.

Process mining contributes to this approach by utilising data from existing tools – such as CRM, ERP or business applications – to analyse how processes actually operate and identify potential improvements. The aim is to select the projects offering the greatest potential return on investment before committing resources.

For senior management, the AI strategy must not become a series of isolated experiments. It must form part of a coherent plan, with clear priorities, measurable outcomes and appropriate governance.
Robby Melin

Robby Melin, Sales Director, NSI Luxembourg

AskNSI: learning from our own experiences to better support our clients

 AskNSI is a prime example of this approach. Available in Microsoft Teams, this assistant helps staff find answers on internal procedures, rules to follow and steps to take. It aims to simplify access to information and make internal support more efficient, particularly in a context where new staff members are regularly joining the company.

Beyond this application, AskNSI serves as a real-world learning environment for testing the mechanisms required to integrate AI agents into an information system: access to knowledge, connection to internal tools, respect for users’ rights and monitoring the quality of responses.

Developed using IBM Watsonx Orchestrate and hosted in NSI’s sovereign cloud, the project enables us to test our technological choices against operational realities. We are also working on developments that will enable staff to access certain personalised information or carry out procedures with the relevant departments, in accordance with their authorisations.

These lessons go far beyond the realm of human resources. The same principles can be applied to IT support, ticketing, sales support or access to business knowledge. The challenge is to integrate AI into work tools so that, rather than merely providing an answer, it can help carry out a task in a controlled manner.

Four conditions for moving into production

Our experience highlights four areas that must be managed simultaneously: knowledge governance, context management, access security and long-term operation.

– Governed knowledge

AskNSI draws on several knowledge bases, including NSI’s internal documents and procedures, as well as publicly available information. The future integration of our SharePoint intranet should also enable us to expand the range of available sources.

Generative Augmented Search (GAS) enables relevant passages to be identified and fed into the model to help it prepare its response. However, simply finding a document is not enough: we must also ensure that it is up to date and authoritative.

If an internal policy changes, the business unit must approve the new version and the technical teams must ensure that the assistant can take it into account. An answer that is correct today may become incorrect tomorrow if its source is no longer valid. AI therefore does not replace information governance: it makes it even more necessary.

- A controlled environment

An agent’s effectiveness also depends on its ability to interpret the information provided to it correctly. The same place may be referred to in different ways, without the AI automatically recognising that it is the same entity.

A common vocabulary, approved by the business teams, and documents enriched with metadata help to minimise these ambiguities. Errors identified during testing thus become a source of continuous improvement.

– Secure access points

Using identity delegation (On-Behalf-Of or OBO), IBM Watsonx Orchestrate enables an agent to call a tool on behalf of the authenticated user, within the limits of that user’s permissions. This mechanism has been implemented and is operational in our AskNSI installation.

An employee who does not have access to a SharePoint document must not be able to view it via the wizard. The application called upon must be able to verify the employee’s identity and enforce their access rights. Traceability of interactions, data protection and control over actions must therefore be built in from the design stage.

– Industrial-scale farming

Users can rate AskNSI’s responses and report issues directly in Teams. But simply gathering feedback is not enough: we need to be able to analyse it and take action. An out-of-date source requires a business-related correction; a search error or technical failure requires a specific analysis.

Confirmed errors must form the basis of a set of tests designed to verify that corrections have been made and to prevent the same problems from recurring after an update. It is also necessary to monitor response times, errors and costs, organise maintenance and plan intervention procedures.

It is this ongoing work that transforms a prototype into a reliable service that staff can rely on every day.
Olivier Baltus

Olivier Baltus, Technical Leader IBM, NSI Luxembourg

Sovereignty: retaining control over data and operations

 NSI’s sovereign cloud enables environments to be hosted in Luxembourg, with local technical support available 24/7. For AskNSI, we rely in particular on Red Hat OpenShift and the IBM watsonx ecosystem.

Our aim is to develop on-demand AI services that enable businesses to benefit from models, inference capabilities, RAG solutions and agents within a controlled environment.

Sovereignty is not limited to the location of data. It also involves retaining control over the architecture, models, data flows and operating conditions. Connections to tools such as Teams or SharePoint must be explicitly defined, taking into account the data being transmitted and the permissions granted.

End-to-end support

 The success of an AI project depends as much on the quality of the models as on their consistent, secure and sustainable integration within the organisation. From defining the strategy to identifying use cases, via data, applications, cybersecurity and operations, every step counts.

NSI draws on its multidisciplinary expertise to support its clients in this process, taking into account their business challenges, regulatory constraints and performance objectives.

AskNSI provides a concrete example of this approach: to ensure that AI is not merely a technological demonstration, but a service that is manageable, secure and useful in everyday life. The aim is to transform the potential of artificial intelligence into measurable and sustainable results.