Why Saint-Gobain Is Building Its Data Foundations Before Deploying Generative AI
As generative AI adoption accelerates in the enterprise, one conviction is spreading across many industrial groups: AI can only deliver value if it's built on data that is prepared, governed and contextualized. At the Tech For Industry Show, Franck Coste, Account Executive (Snowflake) and Simoh-Mohamed Labdoui, Group Head of Data (Saint-Gobain) presented the data strategy the group has put in place to structure this foundation before scaling up AI.
Data, a Strategic Asset at Saint-Gobain
Saint-Gobain, a global leader in sustainable and lightweight construction, designs, manufactures and distributes construction materials through 160,000 employees in 80 countries, under brands such as Isover, Weber and Placo, distributed notably via Point P, Plateforme du Bâtiment and CDO. For Simoh-Mohamed Labdoui, the group's Group Head of Data, data at Saint-Gobain isn't just numbers: it describes the company's know-how — formulations, industrial processes, quality, energy performance, customer uses and market specifics — making it a strategic asset for operational excellence.
A Founding Conviction: Preparing the Data Before AI
The starting point of Saint-Gobain's approach is a strong conviction: artificial intelligence cannot be plugged directly into raw, unprepared data. Solid foundations must be built first.
The Four Pillars of the Data Value Chain
The session details a four-stage value chain:
- Foundations: orchestration, integration, data quality management, monitoring and, above all, access control tools. This last point is seen as critical: an AI connected without control to a tool like ServiceNow could, for example, generate thousands of tickets and uncontrolled compute costs.
- Data access for business teams, with the idea of making data more "conversational" — business teams need answers to their questions ("why is my production line breaking down?"), not just dashboards.
- Business contextualization: the same term can mean different things depending on the organization or site, and AI needs business context to avoid misinterpretation. This contextualization can't rest on IT alone: business teams need to be involved to accelerate it.
- Self-service, through conversational interfaces or augmented analytics, meeting users in the tools they already use rather than forcing them to switch.
Governance Designed to Be Readable by Business Teams
Saint-Gobain stresses that data governance, often seen as complex and costly by business teams, needs to become more operational and easier to understand — through contextualization, clear roles and responsibilities, and data access policies.
Four Success Criteria for Execution
Beyond strategy, Simoh-Mohamed Labdoui identifies four criteria that determine operational success for this approach: simplification (without which adoption fails, and without adoption, no value is created), cost control (reducing TCO to free up resources and prevent business teams from rejecting solutions that are too costly to evolve), the use of mature technologies with AI built in natively, and finally the transformation of the data teams themselves — data engineers, BI engineers, DevOps — whose working methods need to evolve to incorporate AI on a daily basis.
The Partnership with Snowflake
Saint-Gobain has worked with Snowflake for around six years, chosen for its ability to simplify data architecture (relying on data virtualization, limiting duplication) and for a well-controlled total cost of ownership, compared with more complex legacy architectures (Hadoop, Cloudera, Hortonworks). Franck Coste, from Snowflake, illustrates this shift in the relationship to technical expertise with a medical analogy: just as patients now arrive at their doctor's office with a pre-diagnosis already established online, business teams now expect more autonomy and simplicity in how they interact with data.
Concrete Use Cases
A Conversational Assistant for Finance
Named Magnitude, this tool lets finance teams ask questions directly about financial data, building on prior semantic modeling and cataloging work. It aims to give finance teams more autonomy to detect weak signals in their data.
A Group-Wide Data Marketplace
Saint-Gobain has developed an internal data marketplace, designed as a single point of entry to expose the group's data assets: a user based in Germany can discover what sales data exists in Australia or Japan, understand how it's modeled, and request access. The goal is to reduce data access time and associated costs, and to speed up data ownership by business teams themselves — a shift in posture summed up as moving from a "data & AI for all" logic (giving access) to a "data & AI by all" logic (genuinely transforming ways of working).
Architecture Choices to Avoid AI Hallucinations
An important technical point raised during the discussion: where should the semantic model sit — the layer that defines a common business language (customer, contract, stock, plant)? Saint-Gobain chose to build it as close to the data as possible, at the Snowflake level, rather than solely within reporting tools.
Franck Coste clarifies the stakes from a vendor's perspective: historically, semantic layers lived inside BI tools (like Tableau). With the rise of generative AI, these reporting tools are becoming less central to data access. If the semantic model differs between the reporting layer and the generative AI layer, the risk is generating hallucinations or inconsistent answers depending on the channel used. Hence the need to bring the semantic layer closer to the data itself, to guarantee a consistent result regardless of which tool the end user relies on.
Governance, Security and Collaboration: Finding the Right Balance
Asked about data access control, Simoh-Mohamed Labdoui describes a gradual approach rather than governance imposed globally from the outset: defining roles and responsibilities (data stewards), equipping them to reduce their workload, then automating access requests through tools like ServiceNow, Data Galaxy or Collibra. Data classification relies on simple business questions (the impact of a data leak on the business), with possible AI support to automatically detect sensitive data.
On the technical side, Franck Coste notes that Snowflake offers role-based access control (RBAC) that models Saint-Gobain's various entities and geographic zones — a control automatically inherited by every tool that consumes this data, including prompts sent to a generative AI, which can only access data Saint-Gobain has explicitly authorized.
On governance within a global group, Simoh-Mohamed Labdoui advocates a process-oriented approach rather than governance that is either strictly centralized or fully delegated: a single process (for example, a product lifecycle analysis) involves several business functions (supply chain, manufacturing, and so on), which justifies federated governance across the business functions involved, with responsibility distributed according to available maturity and expertise. He also raises the idea that a "true" digital twin isn't limited to IoT, but sits at the level of the process itself.
Finally, on the collaborative dimension, Simoh-Mohamed Labdoui describes a balance to strike between openness and control: broadly open read access to encourage exploration and the emergence of ideas within a given business domain, but more tightly controlled creation rights, to prevent drift while maintaining adoption.
Key Takeaways
The approach presented by Saint-Gobain and Snowflake illustrates a structured take on data in service of AI: solid foundations, governance designed to be readable by business teams, a semantic model positioned as close to the data as possible to limit hallucinations, and a transformation logic that involves technical teams as much as business teams.
Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.
