The assistant implemented by the French group is used by 65,000 employees. Zoom on how its subsidiary Veolia Eau has taken hold of the subject.
At the end of 2023, the Veolia group launched its internal chatgpt. Called Secure GPT, the platform now has nearly 65,000 active users spread over 45 entities, or more than 90% of the multinational white collars. A company which, in total, has 220,000 employees and has 45 billion euros in turnover. 13,000 employees were trained in this technology which brings together 130 internal assistants available on a whole myriad of use cases. The solution implements no less than 19 AI models. Behind the scenes, Veolia uses RAG to manage a dialogue with 280 knowledge bases. It also puts to music 215 agents capable of collaborating between them by relying on the open source frameworks Langgraph and Langchain.
Secure GPT covers the main features of a conversational assistant worthy of the name: dialogue, generation of images, analysis of PDF documents … “For us, within Veolia Eau France, the objective was first to democratize access to internal generative AI”, underlines Meriem Riadi, CIO of the entity in question. To support its employees in the handling of the tool, Veolia Eau France led acculturation sessions throughout 2024, up to putting into practice via prompting. This program was first proposed to the various management committees of the company and then to referents (or “champions”) which were appointed to bring the good word in the region with a view to raising awareness among the teams. In total, 1,500 managers were trained.
Online IA agents
At the end of 2024, Veolia Eau France intended to deepen use cases beyond the basic functionalities of Chatbot, with in particular a focus on the agents. Among these use cases is the automation of the management of consultation files received within the framework of tenders. A project implemented in the fourth quarter of 2024. “The generative AI helps us to analyze these files based on our history of tenders”, explains Meriem Riadi. The stake? Reduce response deadlines to get the contracts (or renew them) with local authorities wishing to delegate the management of the public water service to Veolia. “Secure GPT allows us to save time to focus on the quality of deliverables,” says Meriem Riadi. “Beyond the analysis of the files, we are testing the automatic response generation.” Application code name: Vivaogpt.
Based on the Gemini model of Google, Vivaogpt has made it possible to analyze nearly 150 consultation files since its launch in November 2024. Result: a gain of a working day per file and by business manager. The solution is now deployed to all development teams. “Over a full year, we estimate that more than half of the tenders will be processed via Vivaogpt, more than 300 files. This will correspond to as many day / man saved,” summarizes Meriem Riadi.
“We want to ultimately deploy around twenty assistants within our entity”
A second use case deployed at the end of 2024 concerns customer relations management. Also resting on Secure GPT, it consists of having the chatbot ingest all the customer knowledge bases available within Veolia Eau France: messages exchanged, support documentation, etc. From there, the application, which is based on the Claude 3.5 model of Anthropic, delivers an answer that the advisor can then reread and validate before sending it by email. “Still in the test phase, this tool is an agent articulated around the customer official,” says Meriem Riadi.
Today, Veolia Eau has more than 700 daily connections to Secure GPT for a workforce of 15,000 employees. For the rest, the entity displays a dozen projects in the framing or test phase around AI agents. “We want to ultimately deploy around twenty assistants within our entity. They will go from customer relations around contract management to cases of use specific to IT in particular for application development”, anticipates Meriem Riadi.
At the start, Secure GPT was backed by Microsoft’s Cloud Azure Openai service. “The application was launched in two months. The objective was to quickly lead to a production result,” recalls Meriem Riadi. Veolia then decided to also turn to the AWS Bedrock generative service. “This solution provides us with a whole host of models beyond GPT”, justifies Meriem Riadi. “This is an excellent complement. Knowing that we want to offer our internal users an agnostic environment which allows us to capitalize on a maximum of LLMS on the market, each being able to respond to different use cases.”
A universal platform
Secure GPT intended to be universal and to respond to all the group’s issues in terms of generative AI, it was logical for Veolia to turn to Bedrock. This service opens up access to dozens of LLM in Serverless mode. Beyond the models of anthropic and mistral, it notably incorporates the models of AI21 Labs, Cohere, Deepseek, Luma and Meta and, obviously, Amazon. All accessible via a single API.
In parallel, Veolia also uses the Google cloud. A platform that is used to orchestrate agents, but also as a user interface of Secure GPT. “Nor do we deprive ourselves if necessary to use the Green AI Artificial Intelligence Service of Google Cloud”, specifies Meriem Riadi. This is the case for Vivaogpt. Like Veolia Eau France, each business entity of the Veolia group is in the process of appropriating the Secure GPT application in order to adapt it to its own business uses.
