Microsoft Copilot

Société Générale dropped its in-house SoGPT tool in early 2026 and moved to Microsoft Copilot, a generative AI chatbot built on GPT-4 and free to use. Erwan Simon, CEO of GENIAL, called the resulting build-versus-buy argument a false framing that hides the real question: who owns the business intelligence a company encodes with AI. He reads the SoGPT failure as an execution error, a generic chat interface disconnected from banking operations with no strategy for continuous evolution.

Copilot carries its own limits in that account. It handles generic tasks and struggles on specialized ones such as ERP queries or compliance workflows. Many deployments plateau after the initial pilots: ROI is hard to measure, governance is unsettled, and existing permission structures expose data unintentionally. Simon points to AllianzGPT as the working alternative, a platform orchestrating multiple language models (Azure OpenAI and Anthropic's Claude), connectors to internal systems, and full decision traceability for regulatory compliance.

A World Bank randomized controlled trial put the same free chatbot in a very different setting: a six-week after-school English tutoring program in Benin City, Nigeria, for first-year secondary students. Teachers received three days of training, and prompts were designed around retrieval practice and elaborative interrogation. English scores rose 0.23 standard deviations and overall scores 0.31, at $48 per student for the six weeks.

One product, two readings. The corporate deployment stalls on governance and specialization; the classroom intervention produces a measured gain under infrastructure constraints.