"Today, we can run probabilistic reliability inference on an aircraft. But as a passenger, I'd rather have certainty." That's the line Morgan Zimmermann, Managing Director of the 3DEXPERIENCE brand (Dassault Systèmes) used to sum up the challenge of industrial AI during his Tech For Industry Show session, alongside Christophe Eschenbrenner, DELMIA R&D Generative Experiences Portfolio Manager (Dassault Systèmes). On the agenda: the Industry World Models announced with NVIDIA, the virtual companions AURA, LEO and MARIE, and a deep dive into the shop floor.

The first question mature manufacturers ask themselves: what is the unique knowledge that makes "Renault not Toyota, Airbus not Boeing" — and where is it captured? Most often, it's implicit: hidden in configuration trees, incident histories, documents. Making it explicit requires a great deal of technology. Then comes a tougher question: how to protect it. Because in enterprise AI, "it's not the algorithms that matter, it's the data." Hence the IP lifecycle management introduced by Dassault Systèmes: when a supplier shares data with a customer, they declare whether it can be consumed, integrated — or used for training, and under what rights. Fine-grained traceability makes it possible to know at any time who learned what, from which data, and who holds the resulting intellectual property. Finally, knowledge isn't organized "alphabetically": it's projected onto the twins of products and factories.

Industry World Models combine what can be inferred using large language models with the laws of physics. With a valuable side effect: in realistic virtual worlds, synthetic data can be generated for products that don't yet exist — enabling, for example, predictive maintenance on products that haven't yet left the factory.

Dassault Systèmes talks about "companions," not agents — the craftsman, the friend, the community member. Three profiles for three audiences: AURA for business analysis (risk, projects, regulation), LEO, a certified designer for engineering and industry, and MARIE, a "doctor" in almost every scientific discipline.

The scenario: new customs tariffs hit China and India. Drawing on three twins — cars as configured, the production system, the supply chain — the companion calculates the impact: $3.3 billion. At many manufacturers, answering this question would mobilize "a full team for three weeks."

The companion then instantiates the supply-chain anomaly resolution process by pre-filling the data: the head of procurement identifies priority programs in four clicks, program directors analyze the impact down to the component level (the battery from China…), then dozens of commodity buyers receive optimization tasks with sourced alternatives. The lesson: in a supply chain disruption, "there's one winner and only losers — the winner is whoever moves fastest." And the goal isn't to help executives, but to give every employee the same decision-making power.

Behind the demo, five inseparable building blocks: the reasoning system (sovereign LLMs depending on geography — Mistral or Nemotron in Europe, up to frontier models; guardrails; and shared, stateless AI factories built with NVIDIA, given that a GPU costs €200,000 for an 18-month lifespan); the representation system — the twins, a "map" of industrial operations, like a military command map; the understanding system, which projects data from about twenty systems (ERP, PLM, CRM) onto that map; the engagement system, which connects humans and companions while capitalizing on every exchange; and the action system, which sequences processes at scale. Morgan Zimmermann's warning: your performance will always be limited by whichever system is your slowest.

First use case: at an automaker's methods office, the user defines disc-brake assembly instructions by prompt — where they used to click through menus one by one. The target set by early-adopter clients: cutting the time spent on this step by a factor of 10, while staying in control at every step — "we're not in a black box."

Second case: machining programming guided by LEO, pre-trained on a vast number of machined parts — 30 years of software know-how — to optimize toolpaths and reduce tooling wear, with simulation validation before sending to the machine. Third case: augmented reality on the shop floor (tablet, projector or glasses) to overlay 3D assembly instructions onto the real part and assist quality control — adapted to different usage contexts, day or night.

The DELMIA method: (1) assemble equipment from catalogs — down to a downloadable humanoid robot twin; (2) define the site layout; (3) simulate flows to detect bottlenecks; (4) virtual commissioning to validate real machines; (5) the real-time control tower. AI speeds up creation: starting from a point-cloud scan, the platform recognizes robots, conveyors or forklifts and replaces them with their kinematic models. The final horizon: the self-learning autonomous factory — "dark factories" — trained virtually with NVIDIA and Omron, where the robot learns, in its own context, situations impossible to replicate at scale in the real world.

Asked by the audience, Morgan Zimmermann defended the choice of apparent slowness: plugging an LLM into a system "takes 15 days" but doesn't solve any industrial problem, which starts with security. Hence end-to-end control — GPUs, AI factories, data flows, models — to guarantee sovereignty, intellectual property protection and shared compute power without data leakage. Alongside this: full MCP and agent-to-agent integrations to coexist with third-party ecosystems, end-to-end "Virtual Twin as a Service" offerings, and dedicated startup programs.

Morgan Zimmermann and Christophe Eschenbrenner's session lays out a vision of industrial AI built on three combined pillars: business knowledge made explicit and protected, physics injected into models to move beyond mere probability, and a sovereign, traceable architecture designed to empower every employee, from executives to shop-floor operators.

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