LLMs: A Challenge for the Industry in the Pursuit of Stability

A Breton Company Facing the Challenges of LLMs

In February 2026, a mid-sized company in Brittany, with approximately 1,000 employees, decided to integrate generative artificial intelligence into its operations. This initiative followed a carefully crafted budget established between September and November 2025. Although generative AI was cautiously included in the budget, the massive adoption of so-called agentic tools, capable of autonomously chaining complex tasks, was not anticipated. The goal was to enhance competitiveness, but the company did not foresee the extent of the challenges that this rapid adoption would entail.

A few weeks after the project launch, new versions of the AI model were released, each bringing significant improvements in autonomy. In light of these advancements, management decided to accelerate the process to gain a competitive edge, aiming for delivery in June. However, this decision would prove to be more complex than expected.

The Challenges of Integrating New Versions

Thomas, the project manager, organized the work by allocating specific tasks to staff, software, and AI. Sophie, responsible for quality and compliance, implemented rigorous controls to ensure code safety, traceability, and auditability. Abdel, the technical lead, was tasked with integrating the AI-generated code into the company's security standards.

The team began to use AI in an exploratory manner, discovering its capabilities and adjusting processes accordingly. Operating procedures were drafted, and training sessions were conducted for operators. Sophie even managed to develop online knowledge validation tests thanks to the new generation of LLMs, which would have been too costly before. The communications team promoted the features made possible by AI, highlighting the benefits for customers. Meanwhile, the sales team adjusted the catalog, pricing, and the level of service guaranteed by contract.

However, each new version of the AI required adjustments to safeguards and operating procedures, disrupting the work of operators who had to readjust each time. The prompts no longer produced quite the same results, necessitating constant revisions of procedures.

During the four months of the project, this cycle of updates and adaptations repeated three times, leading to delays and preventing the stabilization of the process. Ultimately, the project was never deployed, illustrating the difficulties of maintaining operational continuity in the face of rapid technological changes.

An Unexpected Bottleneck

This case reveals a shift in the bottleneck: it is no longer the performance of the model that poses a problem, but the time required for human learning and the cumbersome governance mechanisms. The expected productivity gains are absorbed by the need to constantly readjust processes with each new version of the AI.

The company thus finds itself in a delicate situation, caught between the need to keep pace with technological innovations to remain competitive and the necessity of maintaining a structure that guarantees quality for its customers. Despite the skills and determination of Thomas, Sophie, and Abdel, the quality of the processes, which constitutes the company's added value, is incompatible with imposed and uncontrolled changes.

LLMs: Strengths for Language, Limits for Industry

Large language models (LLMs) are particularly effective in areas related to language production and understanding. They are successfully used to automate customer support, synthesize documents, draft commercial offers, train staff, and analyze customer feedback. Companies like Orange, BNP Paribas, HSBC, and Air France have integrated these technologies to enhance their services and internal processes, thereby generating quick value with a short-term return on investment.

However, as soon as one moves beyond the realm of language into industrial processes, the limitations of LLMs become apparent. They struggle to ensure the quality of an end-to-end process, particularly regarding code safety, traceability, auditability, and integration into existing systems. The frequency of new LLM versions, with four major releases of Claude AI between January and May 2026, makes it impossible to stabilize a large-scale process.

A Decentralized Organization as a Solution

Companies that successfully leverage LLMs often share common characteristics: they are agile, decentralized into small autonomous units, and organized around customer satisfaction and value creation. They generate high margins in a limited time frame, and their turnover is often high, as employees are attracted by learning and innovation.

These organizations adopt an emergent, empowering, and bottom-up logic, where teams test, adjust, and retain what works at a sustained pace. However, LLMs are not a one-size-fits-all solution. They excel in language generation but are ill-suited for processes requiring extreme rigor, such as those in industry, law, accounting, or IT. In a forthcoming article, we will explore alternatives to LLMs—more specialized AI technologies that often offer better stability and a more predictable return on investment in industrial environments.