By Constant Ondo
2025 should be, according to an article published by Gartner, the year of AI implementation in companies. And IT departments are at the heart of the project since, according to Gartner again, almost half of CIOs are in charge of steering the AI strategy. They will therefore have to assume responsibility for choices, costs and ROI. Not an easy task, since it is estimated that one in two companies is currently struggling to demonstrate the added value of AI.
It’s safe to assume that these companies should include a large number of industrial firms. This because they have specific requirements for their production activities, which are curiously missing amongst the use-cases modeled by the major consulting firms.
Yet production is the most important internal operation for an industrial company, and AI can be used there both as an assistant for daily work and a transformative force. We therefore felt useful to draw CIOs’ attention on the particular needs of factories in terms of AI solutions, so that they can satisfy all users.
This may lead them to consider the possibility of several AI models cohabiting, as generalist models were not designed for production and will not become so in the near future. This cohabitation does not have to be a problem; on the contrary, it is expected to become widespread in the years to come. It could even be beneficial, reducing complexity, energy requirements and the risk of hallucination.
From chat to automation?
In less than two years, we’ve already gone from experimenting with conversational agents (or chatbots) to creating AI agents. Agents are capable of performing tasks autonomously, deciding for themselves which tools to use and which steps to go through to achieve the desired result. They are automation tools.
Interestingly, AI agents are not using only LLMs. They need other, older forms of AI (like deep learning), and require to be trained on specific use cases.
Why this evolution? Because the results obtained with conversational agents are too often disappointing. And because automation seems to be a good lever to increase AI ROI. So be it. This raises the question of how to automate, and what place will be left for the human at the end of the process. On the question of how to automate, it’s obvious that we can’t approach sales or administration and production in the same way: just as there is OT (Operational Technology) for information systems in the factory and IT (Information Technology) for IS outside production, we need to think in terms of shopfloor AI and front/back office AI. Plants need dedicated AI models.
However, we must avoid falling into the trap of the AI assistant in the office and the AI robot in the factory. It could be tempting to create two sub-projects, with very different scopes and tools. Yet factory teams are multidisciplinary. They need a common tool that can help them generate new ideas, solve everyday problems and automate certain tasks.
Speed, accuracy, confidentiality
On these three points in particular, an AI designed for office tasks can not function properly in the world of production.
Production: a context of urgency
On a production line, every minute lost costs thousands of dollars. It’s impossible to use a conversational mode and refine AI’s answers as you go along. You need to get actionable information from a few keywords or a simple question.
The decision: right first time
A wrong decision made on a machine can have cascading consequences. Choices must therefore be validated before they are implemented. We can’t, as we would for the automatic generation of a text, let the AI produce something and then improve it. Operators don’t want proposals, they want decision-making support based on documented arguments.
The right solution: confidential know-how
Company’s know-how lies in the way it solves a given problem. A company that knows how to set up an injection molding machine better than another, for example, will offer better quality products. A company that knows how to capitalize on best practices will undoubtedly be more profitable. All the knowledge patiently accumulated (about the right settings, maintenance intervals, etc.) is highly valuable, which is why process data is so confidential. It is thus useless trying to retrieve answers from the Internet – they’re not there. And it’s impossible to train generalist models on industrial knowledge – they don’t have access to it.
LLM, RAG and AI agents in a tool designed and trained for industry
To fully understand what distinguishes an AI solution dedicated to the industry from generalist models, it’s important to be clear about the different elements that are LLMs, RAGs and AI agents.
LLMs (Large Language Models) were developed to enable computers to understand and process human language. They are combined with Transformers (Generative Pre-trained Transformers or GPTs), to produce text in a statistical way: words are transformed into tokens and vectorized to “predict” the most likely associations. These tools don’t understand. They decipher, classify and restitute.
The acronym RAG (Retrieval Augmented Generation) refers to a way of generating content in which the AI fetches information from a given context before generating an answer. Previously, LLMs only responded to the data they had been trained on. Their answers could quickly become obsolete.
As we have seen, the AI agent is an autonomous system that creates its own workflows to carry out a given task.
These three elements can be combined in a single, plug&play tool, specially trained for the industry and which automatically updates the company’s knowledge. That’s what we’ve done with PICC.
What’s so special about this solution?
- It is based on a way of structuring information that is common to all industries : problems, solutions, causes, consequences, risks,
- It can operate in RAG mode on large document bases, without any complex integration phase,
- It includes AI agents capable of automatically mapping all the company’s knowledge (explicit and tacit), and capturing new knowledge in all forms (voice, text, image). It thus creates a digital twin of the company’s knowledge, updated in real time.
- It also integrates autonomous agents for the production of industrial objects, such as FMEA, CAPA and 8D.
- It offers two modes of interaction: a “procedural” mode for day-to-day problem-solving, and a conversational mode for continuous improvement,
- Above all, it is capable of automatically correlating human knowledge with production data collected via IIoT systems.
With AI designed and trained for industry, it’s easy to create personas to provide relevant, targeted answers for operators, maintenance technicians, production managers and quality managers. Everyone has access to actionable knowledge based on simple questions.
Achieving the same results with a general-purpose AI will require a colossal amount of customization work, and will explode development costs.
