How a decade of knowledge curation made AI work on the first try

Forty brands, 160 countries, nobody managing knowledge across them

Savencia is a global food group, family-owned, 28,000 employees across 160 countries, nearly 40 brands from Caprice des Dieux to Valrhona. Each brand runs in its own market, but all of them need one thing in common: a subsidiary in Brazil following the same procedures as one in the Czech Republic. That takes governed knowledge at global scale, and it did not exist.

• Eighty-plus systems, each with its own rules. Every business function had built its own. IT delivered what was asked for and moved on. Nobody asked whether any of it connected.
• Nobody owned the knowledge. The role of knowledge manager did not exist. Business units owned their content, but no one was responsible for structuring it, maintaining it, or making it findable across teams.
• PowerPoints passed off as knowledge. A 220-slide training deck would arrive labelled "our policy". Slides in font size 6, full of abbreviations and infographics, are a presentation. They are not a source anyone can search.

Each team had its own system, its own governance. That was it. IT built what was requested, delivered it, and moved on. No one asked whether it made sense alongside what already existed.

From 80 silos to one governed base

Savencia drew a hard line: SharePoint and Teams for collaborative work, Elium for validated, referenced knowledge. Policies, procedures, manuals, directives, anything a subsidiary anywhere has to follow, lives on Elium. The first deployment was market intelligence. It grew into group-wide knowledge management from there.

• A knowledge management function that did not exist before. Anne Thermolle and Vincent Kippelen do not administer a tool. They work with each business function to identify knowledge, structure it and make it findable, challenging what gets submitted and pushing teams from slide decks to structured content.
• Two hundred space administrators across the group. 160 spaces governed by a steering committee, with three system administrators on technical governance. The platform is open to all 28,000 employees and has become, in their words, indispensable and mandatory.
• AI improved content quality before the agent launched. Elium's AI helps rewrite poorly structured pages, detect contradictions and surface duplicates. Better content means better answers, which means more trust, which means more people contribute.

AI helped us unmask problems we couldn't see before. A procedure that contradicts itself halfway through. Content that says the same thing in three different places, but differently each time. The AI detects it immediately.

When AI arrived, the knowledge was ready

HR was first to deploy the AI assistant. The team runs recruitment, annual reviews, onboarding and compliance for 28,000 employees across dozens of countries, on a base built in French and English over several years. The agent was live for 800 HR managers worldwide in weeks, not months. A manager in São Paulo asks in Portuguese and gets an accurate answer grounded in governed French and English procedures. HR ticketing is starting to come down, which was the point.

HR was the first proof, not the last. The same foundation carries market intelligence, group policies and operational knowledge, and each new case deploys faster. Next is an API connection between Savencia's own AI assistant and Elium, so finance, HR and operations query from one interface regardless of where knowledge sits.

Stop with the PowerPoints, stop with the tables in font size 6. Formalise your knowledge as structured content. When we succeed at that, we've already won, even before AI.

Every company is now finding out what its AI can and cannot answer. Savencia had already spent years turning slide decks into structured content, so its own answer came back right the first time.