How L'Oréal Opérations united 22,000 people on a single knowledge platform its AI agents can trust

A million documents and no single place to look

L'Oréal runs several Elium use cases: in marketing, in Research & Innovation, in digital. This one is Opérations, which covers production, sourcing, supply chain, quality and finance. That is 22,000 people across some 40 factories and 150 distribution centres.

• It started clean. Two systems held the procedures and instructions, with SharePoint for everyday files. The split made sense on paper.
• Then every team solved it separately. Each entity bought whichever tool suited its own need, separately configured and separately paid for.
• Covid turned sprawl into chaos. Remote work drove an explosion of Teams sites: one department of 500 people ran up 3,000. The same document lived in several at once.
• Which left no authority. A rule might exist in French, in English, and again as last year's release. Every search returned candidates, not an answer.

Documents get filed, archived, re-synthesised; documents that existed in French and in English get merged into one. Simplifying the corpus reassures the person who has to manage it.

One tool for one job, one place per document

L'Oréal set the rules before it chose the software, then let the rules decide the architecture.

• One tool for one use, one place per document. The durable rules of the business live in Elium, known inside L'Oréal as Kwik. Working files stay in SharePoint, technical documentation in Confluence. Each document is stored once, and everything else links to it.
• The requirement came from users, not from a shortlist. Ninety-five candidate features were gathered from operations-wide surveys, then screened against the tools already in the group. Sixty had a credible answer somewhere, and the gap decided the choice.
• The corpus was cut before it was migrated. Duplicates removed, obsolete material archived, French and English versions merged. The platform inherited a curated set, not a dump. The cleanup was the hard part, not the software.
• Structure follows the process. 113 spaces in about fifteen families by business area, with access rights, versioning and a named owner for every change. New arrivals get five mandatory spaces and choose the rest.
• The agents came last, and answer from those spaces. L'Oréal calls its AI assistants companions. An operations AI companion sits above six sub-domain AI companions, and a question routes to the right domain. The governing document comes back with its sources attached.

Every business area has its own story. The difficulty with supply chain is that we are everywhere. I ran the migration in three waves. First corporate: the group standards, the rules everyone has to follow. Then the zones and divisions publish what is specific to their scope. Then the countries complete it.

From one department to the answer every team works from

Olivier now runs product ownership across the whole of operations. The tutorials sit inside the platform, so colleagues anywhere can learn it on their own.

Ask a question now and the governing document comes back with it, so a rule can be checked rather than taken on trust. That is what the cleanup was for: not a faster search, but an answer with an owner behind it.

The work continues with Elium's engineers: growing the value of the knowledge already captured, managing how it evolves, tracking how the AI agents connected to Elium perform. User feedback closes the loop and enriches the content. A person approves every correction, not the AI.

AI is not the truth. We have to be very clear about that. Our sources of truth are our documents, what is written in our documents. It always calls for a critical eye.

L'Oréal Opérations searched a million documents and got candidates, never an answer. It now has one governed foundation: 8,000 users in the first year, all 22,000 of Opérations today, and AI agents that answer from it with the source attached. Other departments are rolling it out in turn, R&I among them, and the partnership keeps growing.