AI in large organisations: many trials, few services in production
Artificial intelligence expertise responds to a situation most leadership teams know well. Experiments have multiplied: an assistant here, a forecasting model there, generative tool licences handed out to teams. Few of these initiatives have become services used every day, measured and maintained. The gap between demonstration and operation is the real subject.
This gap has known causes. A demonstration works on selected data, with volunteer users, with no security or integration constraints. A production service must answer correctly in cases no one had foreseen, fit into an existing process, respect access rights, keep to a cost per use and make it possible to give an account in the event of error. None of this is settled by the choice of model alone.
Three pressures add to this. The first is regulatory: Regulation (EU) 2024/1689, known as the AI Act, reasons in terms of levels of risk and creates obligations for those who provide AI systems as well as for those who deploy them. The second is economic: executive committees ask what AI brings in, and the answer cannot remain a list of pilots. The third is human: employees already use generative tools, with or without a framework, and the organisation must choose between governing this use and ignoring it.
Our reading: what makes an AI project succeed or fail
Four observations guide our work.
A use case is judged by the process, not by the technology. An AI project succeeds when it changes a precise task, for identified users, with an indicator the business already tracks. A project launched because a technology is available looks for its problem after the fact, and rarely finds it.
Data sets the ceiling. An assistant that queries an outdated document base states false things with confidence. A model trained on poorly defined data reproduces their defects. The quality, freshness and access rights of the data determine what the system will be able to do, whatever model is chosen.
What is not evaluated cannot be put into service. Generative systems are not deterministic: the same question can receive two different answers. Without a representative test set and an acceptance criterion written before development, no one can say whether a new version is better or worse than the previous one.
Adoption is not a final step. A tool that teams bypass produces no value. Users take part in choosing use cases, in testing and in defining what a good answer is. Onboarding and rules of use are prepared at the same time as the system, not after its delivery.
From these observations follows a position: we start from neither a model nor a vendor. Technologies change fast, and the choice made today will be revised. What lasts is an architecture that allows you to change model, an evaluation framework, governance and teams who know how to use them.
How we work: four areas of artificial intelligence expertise
Our approach follows four phases: Frame, Prove, Industrialise, Operate. Framing means choosing the right problems and setting the criteria for success. Proving means checking on real data that the use case lives up to its promise. Industrialising means building the service with its security, cost and oversight requirements. Operating means maintaining and improving it over time. Each area of our offering fits into this sequence.
Deciding where to invest. Before building, you have to choose. Our support for AI strategy identifies use cases, ranks them by value and feasibility, establishes their business case and orders them in a roadmap.
Building assistants and generative applications. Search across a document corpus, drafting assistance, summarising, information extraction: our offering for generative AI in the enterprise covers the design, evaluation and industrialisation of these systems.
Entrusting tasks to agents. An agent does not merely answer: it chains actions in your applications. This autonomy requires permissions, safeguards and human oversight, which we detail on the page dedicated to AI agents.
Governing and ensuring compliance. Inventory of systems, classification by level of risk, roles, approval process: AI governance organises risk control and compliance with the AI Act and the GDPR.
These areas combine. An organisation that is starting out first needs priorities and rules of use. An organisation that already has pilots mainly needs to industrialise and govern. We work on a single area as well as on all of them.
AI is not handled alone: data, cloud and management
An AI system rests on other foundations, and that is often where projects get stuck.
The first is data. Catalogue, quality, access rules, data pipelines: without them, the AI project becomes a data project that does not say its name. Our data expertise addresses these subjects, including putting models into production and monitoring them.
The second is infrastructure. Where models run, which routes data takes, what level of protection applies to sensitive information: these choices fall under our cloud and platforms expertise, particularly for organisations subject to sovereignty requirements.
The third is the management of the project itself. An AI programme involves the business, the IT department, legal and security, with frequent trade-offs. Management Square has its roots in project, programme and portfolio management advisory: we apply to AI the same management discipline as to any other investment, with explicit decision criteria, milestones and trade-offs.
Where to start?
The entry point depends on your situation, not on a standard path.
- You do not yet have clear priorities. Start with framing: which processes, what expected gains, what data available, what risks. The result is a short list of use cases and an investment decision.
- You have pilots that do not make it into production. The subject is industrialisation: evaluation, security, integration, cost per use, operations. We start with a diagnostic of what is blocking.
- Your teams already use AI tools without a framework. The priority is to establish rules of use and an inventory of systems, before an incident forces them on you.
- You need to demonstrate your compliance. The inventory and classification of systems come first. They determine the applicable obligations and the effort to plan for.
In every case, the first conversation serves to qualify your need and to check that our involvement is the right answer. You can describe your context to us via the contact page.