Answers to emails, RGPD audits, brand visual coherence … The FDJ has developed automation pipelines based on generative AI to maximize its productivity.
100% of the winners tried their luck. This rule also applies to generative AI in business: only organizations that move to concrete experimentation manage to generate measurable value. The French Games (FDJ) is one of them. The group has produced smart agents combining automation and generative AI to transform its most time -consuming business processes.
10,000 emails treated with AI per year
This is the case with the most direct return on investment for the FDJ. The Department Provides the FDJ receives 10,000 supplier emails annually on recurring subjects: payment status requests, transmission of various invoices or complaints. The group then developed an automated pipeline which successively mobilizes a software robot and the generative AI: the robot extracts the email from the reception box, an LLM categorizes the email according to predefined business rules. Finally, depending on the message category, a robot recovers the associated data and provides them with the LLM to write an answer. Everything is 100% automated. However, a Fallback system remains in place to redirect non -categorizable responses to a human.
“10,000 emails per year is a lot, but it is not enough to manage to satisfactorily lead a classification model,” explains Nicolas Bouttier, Head of Automation at FDJ. Unlike conventional AI models that require thousands of examples labeled for each category, the generative AI directly includes the instructions formulated in natural language. “Which allows us to go much faster in terms of development, because we have no training time,” said the manager.
The FDJ is based exclusively on generative AI models offered by the company Uipath via an AI Trust Layer. This brick, developed by the multinational, acts as a security wall between company data and external generative AI models. “Using AI Trust Layer, we guaranteed that the data that will be exposed will remain in a private environment, that it will not be found on the Internet,” explains Nicolas Bouttier. Concretely, the architecture makes it possible to exploit the power of GPT or Claude models while retaining the data in the private infrastructure of the company, without transit to the servers of IA suppliers.
Control of contract compliance
Another use case where generative AI excels at FDJ, compliance checking supplier contracts. The group is obliged to audit the contracts of its 300 strategic suppliers annually to verify the presence of eight compulsory clauses: GDPR, insurance, security, etc. “We have a low volume of contract concerned. So, again, leading a classification model, it was quite complicated,” explains Nicolas Bouttier.
The use of the generative AI then makes it possible to explain the clauses to seek directly in natural language, without learning phase. A robot automatically recovers the list of suppliers in the ERP, subjects the PDF contracts to the generative AI which extracts the required information and generates a compliance ratio in Excel format. The report details for each supplier the clause compliance status by clause, allowing purchasing teams to immediately identify contracts requiring an update.
Computer Vision for the control of Visual Branding
Finally and more conventionally, the FDJ uses a computer vision model to control the deployment of its new visual identity in its 27,000 points of sale. The group recently changed its visual identity and had to ensure the correct deployment of the new graphic elements (logos, colors, signage). Rather than carrying out manual controls on site, the FDJ has automated this verification with a computer Vision Microsoft model. “The AI is led to recognize the photos deposited by the installers on our site so as to verify that indeed, the new identity has been deployed,” recalls Nicolas Bouttier. The system automatically analyzes each photo and generates a confidence rate to validate or not visual compliance.
To calculate the concrete king of the generative AI, the FDJ applies a standardized calculation rule: “We first assess the time that an employee devoted to this task devoted, we multiply it by its daily cost, then we deduce total investment in development and in licenses”, details Nicolas Bouttier. The Automation Department of the FDJ, created in 2021, has already demonstrated the effectiveness of this approach with 8,500 hours saved in its 70 traditional automated processes.
