Artificial intelligence put to the test of reality: data, processes and sovereignty according to Gilles Cymbalista
In the ecosystem of artificial intelligence applied to business, where technological promise often precedes real transformation, certain trajectories bring us back to fundamentals. Gilles Cymbalista belongs to this generation trained at a time when data was not manipulated through interfaces, but worked in its materiality, under power and architecture constraints.
Partner and Expert Director in data science and AI at CGI, he comes from a research engineer background rooted in the field. “I am a doctor-engineer, with work conducted in applied research in business. That’s where the real singularity of my background lies.” This dual experience now guides his reading of AI projects.
His “realistic vision of data impact” is rooted in La Redoute, in a model where the customer file constituted the central strategic asset. “This obsession with data quality, I have it in an original way. It’s the foundation.” At a time when computing power was rare, algorithmic optimization was an art form. Today, he notes, the abundance of resources has simplified usage but distanced us from understanding: “we added layers and nuclear power plants because we have the power”. The gain in accessibility has been accompanied by a loss of mastery.
The progressive disappearance of scientific practice
The diagnosis is uncompromising. The industrialization of machine learning tools and the automation of pipelines have transformed the relationship with data. “They take the data, don’t even look at it and put it into the model. They only look at the result.” This practice produces partially functional systems, but it erodes the capacity for interpretation and explainability. However, in regulated sectors such as banking or insurance, the inability to explain an algorithmic decision is prohibitive. “When asked the question: why does this variable have an impact on credit approval? They are not able to answer. We lose the meaning.”
The criticism goes beyond the nostalgia of a generation trained in applied mathematics. It refers to an educational and civilizational issue. Software abstractions have distanced practitioners from the fundamental laws that govern models. “There are laws to respect… a matrix inversion that doesn’t go well, the system doesn’t say so, but it produces a result anyway.” Behind the apparent simplicity of interfaces lies a technical opacity that few organizations truly master.
Hence the call for a refoundation of training. Not an additional program, but systemic reflection involving teachers, businesses and public decision-makers to redefine the fundamental knowledge to be transmitted in the AI era.
The illusion of the technological solution
In large organizations, this loss of control translates into a form of panic. General managements want quick results, even though the structural prerequisites are not met. “They believe that AI is a model that we will plug in. When in fact it’s a complete transformation.” Project failures are not linked to a lack of skills, but to poor assessment of data maturity and necessary organizational transformations.
The question of return on investment becomes central. “Effective AI is expensive. We need to be sure we have an ROI.” This requirement explains his positioning: only propose solutions whose effectiveness is pre-evaluated, to avoid costly experiments without operational impact or meaning. The leader’s role then consists of translating complex concepts into concrete trajectories for business units. “It’s not because it’s a CEO… if you explain it simply to them, you’ve won everything.”
Data as an instrument of strategic legitimization
The experience he conducted in 2020 at a DIY Leader constitutes a textbook case. The analysis of thousands of employee verbatims to build the company’s long-term vision required eight months of data structuring, for four months of analysis. Today, the exercise would be completed in a few weeks. But the main interest lay elsewhere: in the collective legitimization of the strategy. “The restitution was faithful… everyone found themselves in the ideas.”
The creation of the “Pépito” tool, derived from nuggets, designed to detect truly disruptive ideas in the mass, illustrates how algorithmics can reveal what human decision-making processes tend to make invisible. It highlights weak signals carried by an isolated individual.
Training without excluding: another transformation policy
Contrary to certain industrial strategies based on identifying “AI-compatible” employees, Gilles Cymbalista defends an inclusive approach. “We don’t want to leave people on the roadside. Never.” The challenge is not substitution, but the transformation of skills, through learning mechanisms integrated into the tools themselves. The automation of level 1 support tickets, for example, reveals an organizational blind spot: “if level 1 tasks disappear, through what path do we still train level 2 and 3 experts?” For Gilles Cymbalista, the challenge is not to artificially preserve these repetitive activities. It’s about integrating training directly into AI systems, through simulation environments and fictitious cases. Progressive evaluation mechanisms then validate the skill development of employees before going into production. This logic joins a broader conviction: the acceptability of AI depends on how it is introduced. His experiments in call centers show that adoption is immediate when operators perceive a direct benefit, particularly in terms of individual performance.
Sovereignty, costs and hybrid architectures
Sovereignty does not appear here as a political slogan, but as an industrial equation. It is played out on the scale of investments and alliances. “If we try to do it locally in France, we’re not playing in the right court. We’re missing three zeros.” Hence his idea of a European consortium capable of supporting models serving the continent’s players. “Don’t we have an interest in finding a consortium between companies like Thales, Safran… to create a European model for the benefit of Europeans?” The condition is clear: “we need the will and the funding”.
This reading is part of a historical continuity. The current domination of large platforms is primarily based on data mastery. “The master of the world today is Google through data.” The failure of European alternatives then takes on another meaning: that of a missed strategic moment.
However, operational sovereignty goes through hybrid architectures. A protected layer for sensitive data, external models for generic uses. “Large non-sovereign LLMs are more powerful. We have to live with that.” The real challenge becomes flow governance and cost control.
This pragmatic approach leads to reconsidering agentic AI. Far from spectacular demonstrations, it only makes sense when it fits into a clearly identified business process. “An agent on a chatbot is not coherent. It will impact response time.” On the other hand, applied to the optimization of industrial processes, it becomes a strategic lever.
The process twin: towards an AI that transforms without breaking
The innovation he is currently working on illustrates this orientation. Without revealing everything, it consists of automatically reconstructing a digital twin of industrial processes in a sandbox, to test optimization scenarios without modifying the real system. Specialized agents identify bottlenecks, propose AI insertion points, evaluate gains and costs, then replay transformations in the past to measure their impact. “As long as the human hasn’t stamped it, it doesn’t leave the sandbox.”
The interest is twofold: reduce transformation risk and enable progressive industrialization. This vision situates agentic AI not in content generation, but in moving to action. “The LLM is knowledge. The armed wing is the process and the agent.”
Deciding: the last human bastion
The rise of autonomous systems then raises the question of decision-making. In critical domains — health, defense, security — the separation between recommendation and action becomes an ethical architecture. “as in defense, it can evaluate firing solutions, propose them to you, but not fire.” This protective layer constitutes the real space for redefining professions.
On the debate opposing human performance and algorithmic performance, his analysis is nuanced. In the short term, in detection or classification tasks based on massive volumes of data, the machine is structurally advantaged. “Since we have access to a billion identical cases, I doubt that humans can compete.” Human error is born of doubt, where the algorithm applies a probability threshold. But this technical superiority does not mean the disappearance of the human role: it redefines the scope of decision and responsibility.
World Model: strategy as simulation
His interest in Yann LeCun’s work fits into this perspective. The shift from “next word” to “next world” opens the possibility of an AI capable of simulating complete strategic trajectories. “Give me all the scenarios to improve these lines and the world after.” In large organizations, such a system would become a management assistant capable of integrating thousands of weak signals to propose structural transformations. But technological and organizational maturity is not yet achieved.
Relearning to think about transformation
Beyond use cases, Gilles Cymbalista’s reflection brings us back to a fundamental question: how to transform without losing control? And how to preserve the cognitive capabilities necessary to understand the world when tools automate all intermediate operations?
His journey reminds us that AI is not primarily a technology, but a discipline of data, decision-making and responsibility.
