In the telecoms sector, customer experience is increasingly a decisive differentiator. With this in mind, Orange decided to consolidate its leadership in this area, starting with Sosh, its 100% digital brand.
Until then, Sosh offered a very high-quality user journey through its website and app, but one that was exclusively text-based. For a large share of its subscribers, this approach had become conventional, creating a certain gap with the brand’s fundamentally innovative image.
It was with this in mind that, at the end of 2024, Orange France’s management launched the “First” strategic plan. Its aim: to place customer service at the heart of the evolution of the Orange and Sosh brands. To make this vision a reality, the operator made the bold choice to invest heavily in agentic AI and voice-based conversational technology, anticipating the arrival of emerging technologies in order to build an industrial-grade service designed to last.
The Mission Objectives
Converteo was appointed to support Orange’s teams end to end (ideation, development and deployment) in creating this new voice assistant for Sosh. The project’s main challenges were as follows:
- Create an operational virtual advisor, connected directly to Orange’s information system (back office), and able to explain a bill in detail, add an option or guide a customer through taking out a plan
- Adopt a native “voice-to-voice” architecture early, even before it became mainstream on the market, thereby bypassing conventional transitional models (Speech-to-Text / Text-to-Speech) and avoiding a costly future migration
- Develop a tailored evaluation process suited to probabilistic models, whose methodology draws directly on the training given to human agents in contact centres
- Secure the service and make it reliable in real-life conditions (production), relying on a sophisticated technical arsenal: tester bots (virtual customers), an AI-based evaluation system (LLM-as-a-judge) and a bug-hunting programme (bug bounty) strictly dedicated to artificial intelligence.
How the Mission Unfolded
- Phase 1: Exploratory study and selection of the technical architecture
- Phase 2: Development of the multi-agent ecosystem and connection to Orange’s information system
- Phase 3: Testing procedures, set-up of monitoring (observability) and public rollout.
The Review and Results
- From experimentation to industrialisation: The project has moved from prototype (POC) to a robust minimum viable product (MVP). The reliability of responses is now audited across 100% of use cases, going well beyond the stage of a mere technology demonstration.
- Proactive monitoring and agility: The team has a safety net that identifies any regression the night after an update. This gives it the freedom to optimise the product continuously and without fear in a production environment.
- Strengthened security: The vulnerabilities inherent to conversational AI (prompt attacks, etc.) were anticipated, identified and neutralised before the tool was even exposed to end customers.
- Media attention: Sharlie has established itself as one of the very first “voice-to-voice” agents deployed worldwide in a real customer service operation, a feat highlighted on the 8 pm news on TF1.