Agentic AI In Action with Sanofi

With the ambition to become the first AI-powered biopharma company, Sanofi has been investing in AI for years, including the integration of large volumes of data into a unified data mesh and genAI and agentic frameworks. As it tests and launches agentic solutions at scale, we sat down with Andrea Meyer, the company’s VP and Head of Global Supply Chain Strategy, for an on-the-ground view of the reality of implementation, challenges, learnings, and results.

The Business Case

The ambition to become an AI-powered biopharma company was already well established. However, we were still operating with processes and decision models designed for a more stable environment, particularly in supply chain planning and inventory management.

Our journey began with a CEO-driven mandate to shift from reactive to proactive decision-making. The business case centered on data democratization and resilience, enabling real-time insights to anticipate risks and protect patient safety. This required building a unified data mesh and leveraging AI to reduce stockouts and accelerate decisions.

As part of this journey, we partnered with Aily Labs, an AI decision intelligence scaleup, to take an enterprise-wide approach of supply chain resilience, connecting the dots across finance, R&D, procurement, and supply into a single decision layer.

Backed by a top-down, company-wide mandate, we stopped treating AI as an experiment and made it the default, operating model for core routines and operations, with return on AI from day one.

What Sanofi Did

We were starting from a point where we had strong transactional systems and growing analytics, but insights were fragmented and manual; data existed at scale but wasn’t contextualized across processes; supply chain decisions relied on periodic cycles and significant human intervention; and teams had deep expertise but were constrained by manual workload, with new capabilities also needed to operationalize AI at scale. All this shaped our approach, and our path from idea to scale was deliberately incremental, focusing first on building readiness (vs solutions): strengthening data foundations, improving visibility, and establishing governance.

We integrated extensive internal and external data sets into our AI architecture, including the data mesh – a comprehensive repository continually enriched over years, with inputs from diverse ERPs and systems. Establishing foundations like the digital supply chain twin (providing end-to-end visibility of the value chain) took significantly longer than expected, but were critical prerequisites.

Then, building each agent took a few weeks, followed by sprints to augment based on user feedback. We progressed to targeted experimentation, favouring short development cycles and rapid iterations, selecting a small number of high-impact use cases where complexity, value at stake, and decision latency were most acute.

We relied on a hybrid approach, combining external partnerships with internal development, and as solutions matured, the need for governance and lifecycle management led us to the creation of new roles, for example, an AI owner – what Zero100 would call a Translator – who, among other responsibilities, liaises between AI developers and business stakeholders and leads risk mitigation.

How it Went

Challenges:

Key challenges included integrating heterogeneous data sources, managing cultural change for adoption, and balancing speed with risk controls. Ensuring explainability and trust in AI-driven decisions was critical, requiring strong governance and human-in-the-loop safeguards.

Learnings:

My top three would be:

Governance and explainability are non-negotiable for adoption.
Speed matters: short development cycles and quick pruning of low-value features accelerate ROI.
Start with strong data foundations – AI amplifies existing weaknesses.

The Results

We’re currently piloting an Inventory Autonomous Agent with 60 products in our long-tail product portfolio. It coordinates multiple specialized agents to optimize inventory levels: one agent forecasts inventory and identifies out-of-stock risks; another calculates optimal inventory levels; a third assesses opportunities by comparing projected and optimal stock levels; and a fourth oversees implementation. A dedicated agent monitors actions taken by this agent, auditing and ensuring reversibility if needed. An expansion to 200 products (50% of the long-tail portfolio) is planned.

Our conversational IBP Assistant assists senior stakeholders involved in running the IBP process and making decisions. Multiple agents consolidate information from previous meetings – including presentation materials and minutes, as well as structured data and KPIs sourced from the data mesh – to provide contextual understanding and up-to-date analytics. Our long-term vision is for it to automatically prepare documentation, run scenarios, highlight decision points, take minutes, and follow up on agreed actions. Given its executive audience, it was introduced with narrower scope, prioritizing contextual understanding with a roadmap toward greater automation.

On metrics, we assess agent performance using four value drivers: revenue, inventory, cost, and process transformation. Establishing a direct connection between agent activities and traditional financial KPIs is challenging, so our supply chain team collaborated with finance to build standard models that correlate activity with value, based on historical data.

Looking Ahead

Next-generation agents will move from decision support to autonomous execution with embedded risk controls. The integration of predictive and prescriptive capabilities across planning, inventory, and logistics will redefine agility and resilience.

A Piece of Advice

Anchor your roadmap in business value, not technology hype. Start with high-impact pain points, scale iteratively, and ensure governance and change management are built in from day one.

This interview has been edited for length. 