4/10 Do not try to integrate an LLM into an industrial process that requires quality, traceability and stability. This Breton industrial mid-sized company learned this the hard way.
A company recognized for the mastery of its products and services
February 2026. In a mid-sized company (ETI) – around 1000 employees, an industrial activity, a 2026 budget validated after several weeks of discussions between September and November 2025.
L’artificial intelligence generative is there, cautiously. What is not there is the massive adoption of so-called agentic tools: assistants capable of chaining together complex task sequences, for example writing, testing and correcting code. Management wants rapid adoption for competitiveness reasons.
A few weeks later, a new version of the model was released. Then another. In three months, the gains in autonomy are such that management decides: we must accelerate to create a competitive advantage. The project was launched in February, delivery targeted for June. Four months. This is where the story becomes instructive.
A mechanism that seizes with each version
Thomas leads the process. It breaks down the work to be done, allocates tasks to staff, software, AI; the tasks include the controls to be carried out to control end-to-end quality. Sophie, quality and compliance manager, defines and organizes controls: including product code security, traceability, auditability. Abdel, technical referent, must integrate the generated code with in-house security standards so as not to weaken the existing code. Sophie and Abdel integrate AI into their own activity, in discovery mode through trial and error, they discover the capabilities of the AI, evaluate the level of performance and quality and arbitrate accordingly. Around them, we write operating procedures, using the LLM (generative AI model), we train the operators. Thanks to the new generation LLM, Sophie is developing online knowledge validation tests, which would have been too expensive four months ago. The communications team promotes the features made possible by AI, for the benefit of the client. The sales team adjusts the catalog, prices and the level of service guaranteed by contract.
The device is taking shape. After six weeks, it starts to spin. And this is where a new version of AI arrives. It is better – more autonomous, more reliable. Management is pushing for its adoption.
However, this version does not behave exactly like the previous one. Prompts no longer produce quite the same result. The guardrails installed by Sophie need to be adjusted. The operating procedures written by the team must all be reviewed. Abdel analyzes the consequences. The operators, who were just getting their bearings, must train again.
During the four months of the project, this scenario repeats itself three times. Each time, the well-oiled mechanism seizes up and requires a new loop of design, development, quality, adjustment of service levels, updating of procedures, additional training and, sometimes, questioning of an entire section of the architecture. Beyond the delays, we cannot stabilize the process as a whole. Ultimately, the project is never deployed.
The bottleneck has changed in nature
What does this case reveal? That the blocking point has moved. This is no longer model performance – it is progressing faster than expected. It is the human learning time and the cumbersomeness of governance systems. A substantial part of the expected productivity gains is absorbed by the multiplication of cycles (design, development, etc.) with each version upgrade.
The company then finds itself caught in a dilemma: keeping pace with technological changes for competitiveness and maintaining the structure to guarantee quality to its customers. Thomas, Sophie and Abdel lack neither skill nor will: the quality of their process, which provides the added value of the company and its recognition by the market, is not compatible with changes undergone and not decided.
LLMs excel at generating language…
Large language models (LLM) find their full potential in everything relating to language production and comprehension: automation of customer support, documentary synthesis, writing of commercial offers, internal training or analysis of customer feedback.
Large French and European groups are already using them successfully:
- Orange integrates them into its professional offers for support and productivity.
- BNP Paribas and HSBC improve customer relations and internal processes.
- Air France assists its agents with the management of requests and reservations.
These uses generate rapid value, with a short-term return on investment.
… but struggle in an industrial context…
As soon as we leave pure language to enter industrial processes, limits appear. LLMs are poorly suited to guaranteeing the quality of an end-to-end process (code security, traceability, auditability, integration into existing systems, complex operational optimization). Any new version of the LLM weakens the process in which it operates. The frequency of new versions (4 major versions of Claude AI between January and May 2026, excluding Mythos/Fable, and the same for Le Chat and ChatGPT) make it impossible to stabilize a large-scale process.
…and require a decentralized organization
Small businesses and large group entities that perform thanks to LLMs share several characteristics among the following:
- They are agile and decentralized into small autonomous units, close to the field and the customer.
- They are organized around customer satisfaction and value creation, and not around strict adherence to the budget or functional silos.
- They generate high margins over a limited time.
- Turnover can be significant: employees benefit from learning and participating in innovative projects, but the company does not always guarantee a long career path in these entities.
- The logic is emergent, empowering, ascending: the teams test, adjust, keep what works and reject the rest at a very high rate in complete freedom from the rest of the group.
LLMs are not a panacea. They excel at language generation, but are ineffective in processes that need rigor such as industry, law, accounting, IT… In the next article, we will explore alternatives to LLMs – these more specialized AI technologies that often offer better stability and a more predictable return on investment in industrial environments.
