60% of industrial companies worldwide are evaluating AI — yet only 5% have use cases deployed at scale with genuinely measured productivity gains. That finding, drawn from MIT studies, opened this panel at Tech For Industry Show 2026 (Paris Expo Porte de Versailles), moderated by SAP and featuring Sanofi, Forvia and the president of the Elles Bougent association [speakers to be confirmed]. On the agenda: proven use cases, the real barriers to scaling agentic AI, and the accelerators to get there.

Sanofi positions itself as a company "driven by AI" — a course set by its CEO 6–7 years ago that deeply changed the internal culture. The key concept: "snackable AI" — AI usable by everyone, all the time. The result: over 30,000 daily users of the group's AI applications, one third of its 90,000 employees.

On the supply chain side, sales forecasting combines classic statistical forecasts, all internal data and exogenous data — including competitors' stock-outs published by health authorities. The model is approaching "zero-touch" operation, where enrichment was previously done manually.

On agentic AI, Sanofi deliberately advances in stages: staying in proposal mode before granting agents autonomy. Example: a correlator monitors delivery delays across all suppliers of the group's 38 production sites, detecting weak signals at group level. Users keep control, get accustomed, understand — before one day letting the system proactively adjust a lead time or place an order. "It takes time, and that's part of building adherence to the system."

At the automotive tier-one supplier, AI is both old (machine learning for vision- and sound-based quality inspection for over 5 years) and new: the group's data platform counts 20,000 daily users, including 2,000–3,000 on OEE and scrap use cases, with AI companions embedded in dashboards — to the point that prompt libraries now circulate among users.

But the Forvia speaker issues a warning, citing a recent address by Arthur Mensch (Mistral AI) to the French National Assembly: AI productivity for an individual developer reaches 50–60%; at the level of a two-to-three-person team, it drops to 10–15%; at organizational level, it vanishes. User satisfaction is there; the value is not yet — the whole challenge is reorganizing to capture "the euro actually earned."

Forvia focuses on three fronts: planning and scheduling — more reliable predictions, AI-simplified configuration of complex scheduling tools so they can run more frequently, and ultimately centralizing the function outside the plants; maintenance — 5,000 machines across 35 technology categories, increasingly scarce expertise, and a ComEx-sponsored target of cutting MTTR (mean time to repair) by 50%; and data for performance, with a clearly stated frontier: moving from analysis to automatic action, without a human in between.

Wearing both her hats — former automotive industry executive and president of Elles Bougent — the third panelist brings the design angle: where developing a car body side (design, R&D and reliability calculations) required 3-week loops, AI now delivers a first design in 48 to 72 hours. Other opportunities: automatically generating first drafts of OEM tender responses, and IoT-based preventive maintenance. The next frontier: prototype-free design — eliminating physical validation prototypes (radiators, headlights, bumpers) and digitalizing tests as costly as crash tests.

The number-one barrier cited by all: data — but with a precise reading. On the "container" side, the challenge is ontology, the common data dictionary: an LLM can compensate for naming inconsistencies in a POC, but gets lost as you scale. Forvia draws a radical conclusion: its future MES is being built around a "brain" dedicated to managing the ontology of operations, leveraging documented standards such as ISA-95 — on which LLMs have been trained.

A deliberately provocative line from the stage: "don't work on data quality — work on what generates it," meaning the systems (ERP, MES, PLM) and how deeply they are actually used. "AI won't understand that 50% of the process runs in the ERP, but the important part happens in the phone call from Paul to Jacques." At Sanofi, the legacy of over 180 acquisitions forced data aggregation into a data lake (Snowflake), and the migration to SAP S/4HANA — also underway at Forvia — is seen not as a technical project but as an opportunity to harmonize processes. Sanofi is also experimenting with self-healing master data: AI that learns from corrections to automate mass adjustments of safety stocks and lead times, with human verification. And a pragmatic principle: aim for data that is "good enough to proceed" rather than a perfection that prevents you from starting. On the SAP side, the architectural answer is Business Data Cloud: "zero-copy" data sharing to avoid flat-file exports that strip data of its context.

Two further barriers: trust in the tool — at what point does a user accept handing control to an agent? — and confidentiality (patents, sensitive data). The Elles Bougent president adds a rarely discussed angle: non-gendered AI. Only 17% of people working in AI are women, while it has been observed that from 23% women in a data science team, operational effectiveness rises by around 40% — and over 70% of automotive purchase decisions are made by women. Doing without these talents means doing without competitiveness.

Against the reflex of "changing the people at the bottom of the ladder," both manufacturers converge: training starts at the top. At Sanofi, an AI learning program designed with HEC was completed first by the Executive Committee, then cascaded layer by layer down to N-5 managers. At Forvia, reverse mentoring pairs match digital natives with executives — so those who will push AI understand its advantages as well as its limits, and formulate realistic demands.

Forvia's Simplify program, launched by its new CEO, makes AI a pillar of process simplification — daring the question: do we need as many management layers when AI circulates information and accelerates decisions? At Sanofi, an internal AI developer pool called "the Accelerator" operates like a startup: anyone can pitch a business case, delivered within weeks. And 80% of IT support is already handled by an agent. The conclusion fits in one sentence: "It's going to happen anyway. The question is: with or without us."

The final, human accelerator: a culture of discernment. For users — what do I do with what the machine tells me, how do I challenge it? For decision-makers — in a world where information is abundant, it's the quality of information and the judgment applied to it that become critical. The panel's guiding principle: "AI proposes, humans decide" — agents produce scenarios; humans keep strategic arbitration and responsibility.

What is agentic AI? A step beyond predictive AI: agents able to analyze a context, produce scenarios ranked by potential gain, and even trigger actions — with humans remaining in the decision loop ("AI proposes, humans decide").

Why do so few AI projects reach scale? According to the studies cited (MIT), 60% of manufacturers evaluate AI but only 5% measure gains at scale. Main causes discussed: non-harmonized workflows, data without a common ontology, poorly used source systems (ERP, MES, PLM), and individual gains that dilute at organizational level.

Where should data readiness for AI start? With the systems that generate the data: structuring a common ontology (standards like ISA-95), ensuring the real work happens inside the systems, and aiming for data that is "good enough to proceed" rather than perfect.

Should AI agents act autonomously? Gradually: Sanofi's experience argues for staying in "proposal" mode first, giving users time to understand how the system works and build trust, before delegating automatic actions.

This panel was recorded at Tech For Industry Show 2026, the B2B trade show for industrial transformation (Smart Manufacturing, AI, Supply Chain) at Paris Expo Porte de Versailles. Watch the full discussion above, and join us on May 19–20, 2027, Hall 7, for the next edition.