Alan: Agentic Generative AI

Founded in 2016, Alan became France’s first new private health insurer in 30 years and evolved into a technology-first “health companion” serving nearly one million members. In 2024, its AI team, led by Antoine Lizée, launched Mo, an AI medical assistant integrated into Alan’s medical chat. Unlike conventional healthcare AI deployments, Mo communicated directly with patients in real time, while physicians reviewed its responses after, rather than before, they were sent.
The case examines how Alan designed, tested, and deployed Mo in a highly regulated healthcare environment. It explores the system’s multi-agent architecture, retrieval-augmented generation (RAG) approach, layered guardrails, and evaluation framework, as well as the staged, physician-supervised pilot used to assess performance and safety. The pilot achieved strong adoption among eligible patients and received highly positive physician evaluations.
As AI capabilities advance, the case culminates in a strategic dilemma: how far should Alan increase Mo’s autonomy, and how can it balance efficiency, innovation, safety, and trust in a domain where mistakes can have serious consequences?

After the session, participants should be able to:
1. Analyze AI applications at the system level, recognizing the complementary roles of models, retrieval mechanisms, routing, guardrails, evaluation processes, and user interfaces.
2. Explain how retrieval-augmented generation (RAG) and multi-agent architectures can improve the accuracy, transparency, modularity, and scalability of AI systems.
3. Develop evaluation frameworks for AI systems that incorporate not only performance metrics but also safety, user experience, regulatory compliance, and risk management.
4. Distinguish between input-side and output-side guardrails and assess how each contributes to the safe deployment of AI in high-stakes environments.
5. Apply an operational perspective to AI deployment, including workflow design, escalation mechanisms, quality assurance, throughput management, and human oversight.
6. Evaluate the trade-offs between autonomy and trust when deploying AI in regulated settings and determine how governance mechanisms should evolve as evidence and confidence accumulate.
7. Design and prototype an agentic AI solution by combining foundation models, retrieval systems, specialized agents, and user-facing interfaces to address a real-world problem.