"For two years now, if you don't put LLM on a slide, nobody listens." Alstom's touch of humor sets the tone for this Tech For Industry Show panel: beyond the buzz, how do you deploy and industrialize AI in complex industrial environments? Ilaria Michelizzi, Southern Europe Presales Leader (AVEVA), Grégory Bondier, Head of Mobility Data (Alstom) and Salma Bouchiba, Data IA Practice Director (Scalian) answer by combining three perspectives — an industrial software vendor, an engineering consultancy, and a rail manufacturer.

In a complex system like rail transport, the first challenge is aggregating data from internal and external systems, with strong interoperability constraints — every rail project being specific. The value chain: store, prepare for use, then a semantic virtualization layer that will tomorrow make digital twins interoperable.

Deployment at scale runs into three obstacles: sovereignty (heterogeneous industrial standards, regulation, compliance); the data itself (access, volume, complexity, confidentiality, quality, cost); and governance — human (committees, sponsors), technological (tool harmonization) and FinOps (long-term cost monitoring).

Factories evolve over time and don't all share the same maturity level: data needs to be unified, contextualized, and made available through a trusted infrastructure. That's the role of PI System, a globally recognized aggregation platform, complemented by semantic and ontological structuring that speaks to business functions — engineering, operations, maintenance — with self-service indicators and visualization. And because operational resilience now requires sharing beyond the factory walls, the CONNECT platform enables secure, segregated data sharing — including for regulatory tracking.

AI at AVEVA didn't start with LLMs: twenty years of machine learning, notably in predictive maintenance. Three concrete cases shared during the session: Air Liquide, which standardizes its asset naming on the PI System infrastructure and monitors more than 300 sites worldwide with AVEVA Predictive Analytics — detecting deviations months in advance to prepare maintenance and spare parts, across a fleet of more than 40,000 monitored assets; Lamberti, an Italian specialty chemicals company, which used CONNECT to contextualize its data and search for the "golden batch" — achieving a 5% reduction in energy impact with no compromise on productivity, where visualization alone always required a trade-off; and ENCE, a Spanish paper industry leader integrating forestry and biomass, which incorporated uncontrollable external data (drought, water availability): after a 15-day shutdown of its cellulose plant in 2023, anticipation allowed it to avoid shutdowns the following year and cut emissions from its cellulose plants by 15%.

Alstom — rolling stock (trams, metros, high-speed trains, regional trains), signaling and services — manages an installed base of 150,000 vehicles across projects lasting 30 years or more. A structural particularity: trains are operated by third parties, who own the data and set the framework for its use. Sovereignty and interoperability are therefore not options but starting constraints.

Working with Scalian's teams — capable of mobilizing highly specialized profiles, from data engineering to training rail-specific models — Alstom built a platform that removes these obstacles: governance and automatic cataloging of multi-system data, semantic layers for interoperability at client sites, and above all consistent cloud/on-premise deployment — since SaaS isn't accepted by every operator, many of whom rightly require applications to run on their own premises. The result: calculating reliability indicators for every train across all clients without reinventing the wheel for each project, from France to Germany, from Singapore to Montreal.

Scalian (more than 5,500 employees) claims dual technical and industrial expertise: systems and software engineering, operational excellence, and digital solutions (MBSE, digital twins, simulation, AI) — organized into business practices and centers of excellence, such as the SIMU center. Its AI approach follows four steps: strategy, security & compliance, data foundations, then deploying use cases with monitored ROI. With proprietary accelerators such as AURA, a sovereign, plug-and-play agentic platform (on-premise, cloud, hybrid), and references such as Colibot, an embedded AI for guiding an airdropped package, developed for the DGA (French defense procurement agency), trained on a digital twin and executed on an onboard computer — synthetic data from simulation making the algorithms "three times more effective."

The problem Alstom faces: integrating its data into "the digital twin in France, in Germany, in Montreal" without redeveloping it each time — and, symmetrically, allowing operators to integrate an Alstom twin, a signaling twin, an infrastructure twin. Hence the challenge of the interoperable digital twin, addressed technically through semantic layers and organizationally through shared governance: Alstom is working with the rail ecosystem on a European data-sharing dataspace.

The panel's synthesis: a genuinely interoperable digital twin bypasses the three initial obstacles — sovereignty, because the data never leaves; governance, because it creates a shared reference framework that forces monitoring; and data availability, because it can generate and contextualize it. And AI, through LLMs, becomes a dialogue interface with one's own sovereign twin — trained and enriched on its own data, not someone else's. With one condition reiterated by Salma Bouchiba, of Scalian: without business knowledge — use cases, interoperability, compliance — none of this is achievable.

This panel shows how three complementary players — vendor, integrator, manufacturer — converge on the same conclusion: sovereign AI can only be deployed at scale through genuinely interoperable digital twins, capable of simultaneously bypassing sovereignty, governance and data-availability challenges, provided they remain grounded in solid business knowledge.

Watch the full panel, along with the other Tech For Industry Show replays, on our dedicated page.