AI Project Failures: The Wrong Tool in Business

A Majority of AI Projects Fail to Scale

In the realm of enterprise artificial intelligence, a troubling reality persists: approximately 80% of AI projects fail to scale. This figure is often cited in discussions and conferences, highlighting a recurring issue. The reasons put forward include resistance to change, data quality, and a lack of skills. However, a more fundamental and less discussed cause may be behind these failures: the use of the wrong tool.

The Limitations of Large Language Models

In many cases, particularly in the insurance sector, companies find themselves repeating the same scenario. They launch proof of concepts (POCs) with large language models such as ChatGPT, Mistral, or Gemini. These projects engage teams for months but often end up failing in production. Reasons include a high error rate, skyrocketing costs, and security policies that prevent data transfer to the cloud infrastructures of these giants.

Large language models (LLMs) are indeed impressive for tasks like content generation, document synthesis, or marketing. However, their universal design, based on billions of parameters for millions of users, makes them ill-suited to the specific needs of an organization.

For tasks such as automating claims management or processing invoices, accuracy and compliance are crucial. LLMs often fail where Small Language Models (SLMs), specialized and trained on the company's internal data, excel with accuracy rates reaching up to 99%.

Take the example of a claims declaration. Traditionally, the insured contacts an advisor, receives an email, sends back the necessary documents, and the processing takes several weeks. With an SLM, this process is instantaneous: the agent sends the appropriate form and verifies the documents in real-time, a task impossible for a generalist LLM.

Data Sovereignty, an Unavoidable Requirement

In sectors subject to strict regulations, the priority for IT leaders is not only the efficiency of a tool but also the assurance that data remains internal.

Cloud-hosted LLMs cannot guarantee this security. In contrast, an on-premises SLM can, due to its low resource consumption. These specialized models are several hundred times less resource-intensive than a large model and can operate on the company's servers without requiring data transfer outside.

For companies subject to GDPR, this autonomy is not just a selling point but a fundamental requirement. Large enterprises have understood this: before discussing use cases or return on investment, they demand certifications such as ISO 27001 or HDS to ensure that no sensitive data leaves their infrastructure. This technological lightness allows compliance with these requirements without compromising performance.

Cost Control, a Strategic Advantage

Another often underestimated advantage of SLMs is cost predictability, an essential criterion for financial departments.

LLMs, accessed via APIs, charge based on tokens, a unit of measurement related to the size of the texts processed. This usage-based billing is difficult to anticipate and can blow the budget between the testing phase and production deployment.

In contrast, an on-premises SLM escapes this logic. Its cost depends on the company's hardware infrastructure, offering total budget predictability. This stability is crucial in a sector where rigorous cost management is as important as risk management.

Many AI projects have yet to reach their full potential, not due to poor direction but due to a poor choice of tool. The universal solutions offered do not meet the needs for specialization, sovereignty, and predictability in regulated sectors. The companies that succeed today are those that have opted for models tailored to their specific constraints, a lesson learned after several years of unsuccessful experimentation.