• A multi-agent orchestration system developed by the startup ILLUIN Technology enables Sharlie to autonomously and seamlessly handle a wide range of customer requests.
• To test the system on a large scale, Orange has created simulation environments and virtual customers.
In March 2026, Orange announced the launch of Sharlie, Sosh’s voice assistant. This innovation stands out in the market thanks to its uniquely designed interface, the result of a year and a half of research and development and technological partnerships. Available to customers, this agent-based AI tool—known as “speech-to-speech”—can automatically handle customer service tasks (such as plan sign-ups and contracts) and help customers understand their bills, purchase a phone, or even sign up for new options.
The interface operates directly from voice to voice
Voice responses to spoken questions
“Here, the customer’s voice is handled directly. Unlike traditional speech-to-text architecture, which converts speech to text and then converts the model’s textual response back to speech, the interface operates directly from voice to voice,” explains Mauricio Bezerra Cardoso, a data scientist at Orange Innovation, who contributed to this project with Orange France. Available 24/7, Sharlie is integrated with Orange’s customer intelligence tools and guarantees personalized, fast, and reliable responses.
A well-oiled multi-agent orchestration
Behind Sharlie’s interface, several AI agents process requests. “We’re working with the French startup ILLUIN Technology for agent orchestration. Their Dialogue platform allows us to create different agents and orchestrate them, so we know which agent a customer’s request will be routed to.” Thus, if a customer wants to add an option to a plan, a specific agent will handle their request. The structure is designed as a funnel: “There is a reception agent who will clarify the request; then, if the customer is logged in, an agent will select the relevant contract, and the request will be routed to the agent specialized in the customer’s intent.” The goal is to never expose this mechanism to the user: “We want the customer to feel like they’re talking to a single entity.” The team is working to further improve this architecture with this priority in mind.
Virtual customers to test the service at scale
During the development and testing phases, a dedicated simulation environment allowed the teams to develop Sharlie without relying on Orange’s actual information system. “The idea was to provide simulations of the information system so that developers could understand how Orange’s information systems would respond to each customer request.” Rather than enlisting real customers for large-scale testing, the team deployed another generative model that provided virtual customers. “We built another LLM, another generative model that poses as a customer and interfaces with Sharlie to perform non-regression tests and test new features.”
For customers, the experience is enhanced. Sharlie guides them through the sales process to help them choose the plan best suited to their needs. Subscribing to a plan requires online validation by the customer, who retains the standard right of withdrawal in such cases. “However, AI agents are fully capable of independently adding new options to an existing contract,” emphasizes Mauricio Bezerra Cardoso.
This text has been translated by an artificial intelligence.
Read more :
Orange France lance deux nouveaux services d’intelligence artificielle pour enrichir sa relation client (press release, French)
Mauricio Bezerra Cardoso
