The fragmentation of LLMs delays companies’ AI projects and significantly impacts the return on investment of their initiatives.
The publication last year of the third stage of the national AI strategy in France reaffirmed the country’s ambition to fully enter the era ofartificial intelligencesupporting its adoption in businesses and public services. While this roadmap provides certain dynamism, it also highlights the concrete challenges faced by companies trying to transform this technological promise into added value. At the heart of their concerns, an essential question remains: how to guarantee a tangible return on investment (ROI) in the face of the growing complexity of the AI ecosystem?
Despite the immense potential of AI, companies often find themselves faced with a diversity of systems, each more promising than the last. Should we and how should we choose? Moreover, when each seems to require specific, long and costly developments, hindering fluidity and scalability.
The promise of AI meets the skills gap
The first pitfall is not technological, but human. The enthusiasm for AI is palpable, but this wave of innovation is coming up against a glaring skills gap.
A recent Deloitte study highlights this dichotomy: if the adoption of AI has accelerated significantly in 2025, the deficit of qualified profiles capable of understanding, deploying and governing these technologies remains a major challenge. Without sufficiently competent teams, the promise of AI remains a dead letter, or worse, generates unoptimized investments.
The “Wild West” of LLMs: Fragmentation and hidden costs
This skills gap is exacerbated by the fragmentation of the AI landscape. Each major language model (LLM) comes with its own specifics, APIs, and unique integration requirements. This lack of common protocols and interfaces transforms the governance and orchestration of this ecosystem into a real headache.
As a result, companies are forced into specific, time-consuming and costly developments for each new AI brick involved. Integration with existing business tools becomes complex, fragile and difficult to maintain. The observation is clear: this lack of standardization acts as a real burden and considerably dilutes the potential return on investment of AI initiatives.
The problem of trust
When each AI tool deploys its own “language”, its own inherent biases and opaque decision-making mechanisms, how can users, CIOs and regulatory bodies fully trust them?
The effectiveness of AI directly depends on its predictability and its ability to be understood and audited. A solution that generates results that are unexplainable or inconsistent with other systems creates more skepticism than value. Trust is the fuel for mass adoption, and it can only come from better readability and standardization of interactions.
These challenges of skills, fragmentation and trust should not slow down our momentum, but encourage us to structure our approach. The potential of conversational and autonomous agents, capable of orchestrating different AI bricks and interacting in a more natural way, opens a promising path. A framework would not only encourage the development and smooth use of AI in business, but also maximize its capacity for innovation and competitiveness.
