At the Gartner Data & Analytics Summit last week, there was no shortage of AI theatre. Every second presentation seemed to promise autonomous systems, predictive operations and self-improving processes. The language was familiar by now. What stood out instead was a session from KME, largely because it avoided most of that vocabulary altogether.
KME spoke about industrial AI in the context of an actual manufacturing plant rather than a conceptual architecture diagram. That sounds like a small distinction, but it changes the conversation considerably. Inside industrial organisations, the difficult work is rarely the model itself. It is the operational environment surrounding it.
Operational procedure
KME operates complex copper production processes at its Fornaci di Barga facility. The plant generates continuous streams of operational data across machines, production stages and control systems. None of this is unusual in modern manufacturing. Many factories already produce more data than they can realistically use. The problem is that most of it exists in disconnected systems built at different times, often for different operational purposes.
That fragmentation creates an awkward dependency on human experience. Operators compensate for missing context through instinct and accumulated local knowledge. Engineers build informal workarounds between systems that were never designed to communicate properly. Management presentations often describe these environments as “digitally transformed” long before the underlying operational coherence exists.
The company did not frame its challenge as a shortage of AI capability. The challenge was establishing a trusted operational picture across the plant in the first place.
Data
One of the more important observations from the session was also one of the least fashionable. Good AI depends on good operational data. Most people nod when they hear that sentence, but industrial environments give the statement sharper consequences than many software sectors do.
If Netflix recommends the wrong film, nothing material happens. If a factory operator loses confidence in process data tied to quality or throughput decisions, the system itself quickly becomes optional. Production environments are unforgiving that way. Once trust erodes, people revert to manual judgement remarkably quickly.
KME seemed to understand this early. Much of the work described in the session was not centred on AI models at all. It was about consistency, accessibility and context. Machine data needed to become reliable enough for operational teams to use without second-guessing its validity every few hours.
That led to the development of what KME described as its “KME 5.0” data platform. The phrase itself is less interesting than the operational intent behind it. The company was effectively trying to create a common layer across systems that had evolved independently over many years.
Integration
The integration work sounded particularly familiar to anyone who has spent time around industrial programmes. AI discussions often focus on analytics layers because they are easier to demonstrate publicly. Integration is slower, more political and much harder to present elegantly on a conference stage.
Factories accumulate technology unevenly. Different production lines operate on different upgrade cycles. Some equipment remains operational for decades because replacing it interrupts production economics that still work perfectly well. Vendors change. Interfaces change. Naming conventions change. Meanwhile, leadership teams continue asking why “the data” cannot simply be unified.
KME worked with SAS and Alleantia to create an OT/IT integration layer across these heterogeneous systems. Again, this is not glamorous work. But it is usually the point where industrial AI projects either stabilise or quietly stall.
Discipline
There is also a less discussed trade-off here. Greater operational visibility often exposes inconsistencies that organisations have learned to tolerate informally. Once data becomes transparent across functions, disagreements surface more visibly as well. Different teams discover they have been measuring the same process differently for years. Local optimisation habits suddenly become visible at system level. Operational politics enters the conversation whether anyone planned for it or not.
What KME described afterwards was not a dramatic leap into autonomous manufacturing. It was a gradual increase in operational maturity. First came visibility through real-time dashboards and traceability improvements. Then correlations between process conditions and production outcomes became easier to identify. Eventually, predictive maintenance and optimisation work became possible because the underlying operational data environment had become more stable.
That progression matters because it runs counter to how many AI programmes are currently marketed. There is often pressure to demonstrate advanced AI capability early because it attracts attention internally. The less visible foundational work receives less organisational patience, despite being the part that determines whether anything scales later.
KME’s sequencing felt more disciplined than ambitious. In practice, that may be why it sounded credible.
Culture
The most useful part of the session, though, was probably the discussion around culture. Industrial organisations tend to trust systems that survive operational pressure, not executive messaging. People who keep plants running are usually sceptical for good reason. They have seen management fashions come and go before.
KME acknowledged this openly. Adoption happened gradually through small groups of operational champions rather than through large transformation rhetoric. That approach is slower. It is also often more durable.
One comment from the session stayed with me afterwards: “The data platform becomes the n+1 production system.”
That observation captures something many organisations are only beginning to confront. Data infrastructure is no longer just supporting production indirectly. Increasingly, it shapes how production decisions themselves are made. That changes organisational power structures in subtle ways. Operational expertise still matters enormously, but it now competes and collaborates with increasingly systematised forms of visibility and optimisation.
Smart people disagree about where this ultimately leads. Some see industrial AI strengthening human operational judgement. Others worry that over-reliance on system-level optimisation gradually weakens local expertise that factories still depend on during abnormal conditions. Both arguments contain some truth.
What I heard was a manufacturing company working through the practical realities of operational AI without pretending the difficult parts had disappeared. Frankly, that made the session more useful than most.
