Generative and Agentic AI on the Shop Floor: Velotic's SCADA and MES Demo
Combining artificial intelligence with long-accumulated industrial know-how: that's the promise behind Laurent d'Izarny-Gargas, Regional Commercial Leader (Velotic), a new brand presented for the first time in France at the Tech For Industry Show.
Velotic: A Young Brand, Decades of Industrial Experience
Officially created on March 1, 2026, Velotic was born from the strategic combination of Proficy, the industrial software brand originating from General Electric, and two well-known PTC products: Kepware and ThingWorx. Despite its youth as a brand, the company starts out with around 1,200 employees and nearly $350 million in annual revenue, with the ambition of becoming a leading independent player in industrial data and manufacturing.
The Velotic portfolio is structured across three layers: at the base, Kepware handles data aggregation, extraction and normalization; at the operational level, the Proficy Smart Factory MES, a data historian, and two well-known market SCADA systems (Simplicity and iFIX) make it possible to use this data to improve performance; finally, ThingWorx forms the upper intelligence layer — presented no longer as a simple "IoT platform" but as a "data IoT platform," designed to host artificial intelligence and agents in service of business applications, notably industrial asset management.
A Starting Observation: Often Unrealistic Expectations Around AI
Laurent d'Izarny-Gargas opens his presentation with a figure: according to a study cited during the session, 71% of IT and technology directors believe their senior management has unrealistic expectations about AI. He attributes this to the proliferation of pilots and proofs of concept, without the move to industrial deployment — which demands a much higher level of stability — keeping pace, delaying the demonstration of ROI. A broader shift is emerging, according to the speaker, toward greater selectivity in choosing use cases to industrialize, rather than multiplying scattered pilots.
Five Recurring Obstacles to AI Adoption in Factories
The session details the main obstacles observed:
- Data: many factories rely on older PLCs, whose protocols are poorly documented or barely maintained, making data extraction complex.
- ROI: the need to define clear indicators from the outset, and to measure not just benefits but also integration and operating costs, as AI models become increasingly expensive to run.
- Adoption: a high-performing technical solution isn't enough without training, appropriate ergonomics, and the trust of operators and engineers — change management remains decisive.
- Security, with a particular focus on the risk of "data poisoning": corrupted production data (bad inspection reports, incorrect quality measurements) can influence an AI agent's decisions, raising a data-compliance issue as much as a network-access one.
- IT/OT environment heterogeneity: older machines, outdated operating systems. Velotic claims an ambition to communicate with all existing systems rather than forcing their replacement.
Three Maturity Tiers of Industrial AI
Laurent d'Izarny-Gargas distinguishes three technological layers, noting that a fourth could emerge as the field evolves rapidly:
- Machine learning, deployed for around fifteen years, initially for predictive maintenance, considerably strengthened by the arrival of neural networks.
- Generative AI, now widely used to retrieve information that was previously hard to access (technical manuals, reports, SCADA event logs) and enable natural-language interaction with industrial systems.
- Agentic AI, which encapsulates the two previous levels within systems capable of choosing the best method to solve a given problem on their own — still at an early stage in industry, with the exception of certain "dark factories" already operational in Asia.
Concrete Examples, from Machine Learning to Agentic AI
Machine learning was first used in steelmaking (furnace optimization), then in process chemistry (continuously maintaining the best correlation between parameters, or "centerlining"). A more recent and less expected example: a Nissan plant near Nashville, in the United States, uses an analytics tool to measure robot wear based on force torques measured at different points, making it possible to anticipate gearbox wear. On the generative AI side, a now-common use case lets an operator quickly find, within a technical manual that can run to 2,000 pages, the solution to a problem identified by predictive maintenance.
Demo 1: An AI Layer on Top of SCADA
Faced with SCADA systems capable of generating dozens, even hundreds, of alarms per minute during an incident, Velotic presents an additional analysis layer — named Proficy Studio — positioned on top of the control system, without replacing it.
The demo illustrates several capabilities: the AI automatically organizes raw, unstructured data from multiple SCADA systems (potentially representing several plants) into a four-level hierarchy, inferring the plant's structure on its own from the collected parameters — a proposal to be validated and edited by teams before going live. The tool also lets users view active alarms, add comments and resolution notes visible to SCADA operators, compare an alarm's history against different indicators on a chart, and identify correlations between alarms — for example, a pressure issue occurring 507 times alongside a low level on a given piece of equipment. Once the plant's structure is identified, the AI can automatically generate a control screen.
Demo 2: Conversing with the MES in Natural Language
The second demo covers the interaction between a production manager and their MES (Manufacturing Execution System), a system handling real-time production tracking, distributing instructions to operators, collecting quality data, and interfacing with the ERP.
Connected to an AI assistant, the MES answers natural-language queries: listing the last ten production stoppages with their causes (with immediate insights on ongoing stoppages, identified overheating cases, or missing descriptions for certain events); ranking the 100 most significant machine stoppages by importance in a Pareto analysis, revealing that the top five causes — overheating, startup issues, quality stoppage — account for 80 to 82% of interruptions; or identifying the best- and worst-performing machines. The speaker notes in passing a deeper issue revealed by this analysis: the absence of a cause description for a large majority of events, symptomatic of a broader data quality issue.
Four Recommendations for Deploying AI in Production
Laurent d'Izarny-Gargas concludes with practical recommendations: give teams room to experiment while keeping control, to avoid ending up with thousands of unmanageable AI pilots; always start from a concrete business problem to solve rather than looking for a use case to justify using AI; treat data quality as the basic condition for any reliable AI; and never overlook the human factor, given the need to build a genuine culture of using these tools within the company.
Velotic's Vision: Combining Artificial Intelligence and Authentic Intelligence
The session closes with Velotic's central message: while industrial assets hold a lot of value, a company's greatest asset, according to the speaker, remains "authentic intelligence" — the experience accumulated over time by engineers and operators. The stated ambition is to combine this human intelligence with artificial intelligence, following an analogy borrowed from aviation: a "cruise" mode, where models deemed reliable enough can run with a high degree of automation (predictive maintenance, automatic scheduling adjustments), and a "takeoff/landing" mode, where humans stay in control, assisted but not replaced by AI.
Key Takeaways
This session illustrates a pragmatic approach to industrial AI: starting from concrete shop-floor pain points — too many SCADA alarms, poorly documented production stoppages — rather than from AI itself, while keeping humans at the center of the system through the notion of authentic intelligence.
Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.
