In industrial AI, a wrong answer isn't a bad chatbot reply, it can mean a shutdown, a market collapse, or worse. That distinction is why Juergen Weichenberger, who has built AI systems since the 1990s and now leads data and AI work at EY after running AI at Schneider Electric, argues the industrial sector remains the least AI-penetrated part of the economy despite having some of the highest value at stake. His answer for why isn't a technology problem. It's a trust problem, and a design problem, built on decades of hard lessons about what happens when optimization goes further than physics allows.

Who is Juergen Weichenberger?

Juergen Weichenberger has been building AI systems since the 1990s, long before the field had mainstream attention, and long before it involved chatbots at all. His early research included building a barrier-free parking garage system in 1996 using number-plate recognition, work that would now be considered a routine student exercise but required real innovation at the time, and a full digital twin of an entire airport, modeling every moving piece from passengers to aircraft to security. That early emphasis on real-world, high-stakes applications led him into industrial AI, where he has worked on problems ranging from improving oil and gas distillation column efficiency by half a percent, worth six billion dollars annually at scale, to critical national infrastructure like electric and gas grids. He led AI at Schneider Electric and is now a data and AI partner at EY.

Why does the industrial sector lag behind in AI adoption?

Weichenberger points to a specific cultural gap: engineers are trained not to trust black boxes, and AI practitioners are often not good at explaining what their systems are actually doing. Marketing, finance, and insurance have all moved faster on AI adoption. Industrial environments haven't, not because the value isn't there, but because the two groups responsible for building AI into these systems don't naturally speak the same language or share the same risk tolerance.

What does an AI project in critical infrastructure actually look like?

Weichenberger describes current work on electric and gas grids as a problem of runaway complexity. Aging infrastructure now faces rising demand from data centers, EVs, solar, and battery storage, none of which the original systems were designed to handle. The real challenge isn't a shortage of options, it's an overwhelming surplus: five million possible configurations where a human decision-maker ultimately has to approve one. Load balancing has grown just as complex, shifting from a simple one-directional flow from a power station to a system with tens of thousands of individual consumers feeding power back into the grid simultaneously. The first job of AI in this context, according to Weichenberger, is making that complexity human-comprehensible again, not replacing the decision-maker, but narrowing millions of options down to a handful a person can actually evaluate and approve.

Why does optimized AI sometimes break the very system it's optimizing?

This is where Weichenberger's industrial experience diverges sharply from how AI is typically discussed. When you ask an AI model to find the most optimal solution, it will genuinely find the mathematical peak, but that peak is often razor thin, with almost no margin for error. He compares it directly to spacecraft engineering: NASA doesn't build the maximally optimized spaceship, because the margin for error at that theoretical peak is too small to survive contact with reality. He cites a conversation with Apollo and Space Shuttle flight director Gene Kranz, whose famous principle, failure is not an option, captures exactly the discipline industrial AI has to be built around. Left unconstrained, AI has a tendency to violate the practical boundaries of physical systems: gravity doesn't negotiate, and neither does the load capacity of a downed transformer.

The same dynamic shows up commercially, not just physically. Weichenberger describes a real case where AI was told to optimize product throughput without constraints, and it succeeded, flooding the market with enough supply to collapse the product's price and the company's margins along with it. His clients now routinely specify not just what to optimize, but where optimization should stop, because the theoretical best solution on paper is often several steps removed from the point of highest actual profit.

Why has predictive maintenance failed in nearly every project he's seen?

Weichenberger is blunt that predictive maintenance, one of the most repeated promises in industrial AI over the last seven years, has not delivered value in a single project he's personally observed. His example: a fertilizer plant chief engineer who told him plainly that he can already read his gauges and knows when equipment starts to fail. More importantly, the plant's ammonia production system was built with inherent redundancy, three pumps where only two are needed to run at full capacity, so when one pump showed wear, the team simply switched to the standby and continued operating without any production loss. They didn't even rush to fix the failing pump, because there was no urgency: they scheduled the repair for the next planned shutdown. Predicting the failure earlier would have changed nothing, because the engineers already had a working system for managing it. The lesson Weichenberger draws is a simple one: a technically impressive AI capability that doesn't address an actual unmet need creates zero value, no matter how compelling the pitch sounds in a conference room.

What actually moves the needle in industrial AI?

After watching enough so-called game-changing projects fail to produce real ROI, Weichenberger and his team narrowed the value drivers down to four fundamentals: yield, the conversion ratio between raw material input and finished output; energy, the input required to produce that output; throughput, how much can be produced at the same yield and energy level; and quality, which determines what a company can ultimately charge. He argues these four factors consistently outperform flashier AI use cases, because the flashier use cases are often working around problems the underlying process has already solved through redundancy, planning, or existing operational discipline.

What happened when 150 consultant-recommended use cases met actual operations staff?

Weichenberger shares a striking example from a project in South Africa, where a chief digital officer handed him a list of 150 AI use cases identified by a major consulting firm and asked him to implement five. When he asked whether those use cases had been validated with operations, the answer was no, the recommendation had come from a trusted firm and that was considered sufficient. After running the list past actual operations staff, 149 of the 150 use cases were eliminated, leaving one. When Weichenberger asked operations directly what problems actually bothered them, they produced ten specific, longstanding pain points, including one asking for just five more minutes of uptime per day on a piece of mining equipment. Multiplied across fifty machines running continuously, that single small ask translated into roughly 300 million dollars in additional annual production. The gap between what gets recommended in a boardroom and what actually creates value on the floor, in his experience, is almost always closed by talking directly to the people doing the work.

What's the most underrated skill in AI right now?

Weichenberger's closing point is about the people building these systems, not just the systems themselves. He describes working with graduates from top technical programs, MIT, Stanford, Berkeley, Caltech, Oxford, Cambridge, who are trained to ask for the data first and run it through a standard process, without first asking for context about the problem they're solving. He argues context engineering, understanding a business process deeply enough to know what a dataset actually represents before applying a standard technique to it, is dramatically underrated relative to model selection or architecture. His advice for anyone deploying AI right now, in a factory, a hospital, or any operational environment, is simple: sit down with the people doing the work, listen to their actual burning problems, and let that conversation shape the project before anything else does.

FAQ

Why is the industrial sector the least AI-penetrated despite the highest value at stake? Because of a persistent trust gap: engineers are trained to distrust black-box systems, and AI practitioners often struggle to explain their systems in terms industrial teams can verify against physical reality.

Why does predictive maintenance often fail to deliver value? Because many industrial systems are already built with redundancy and planned maintenance schedules that make predicting a failure earlier irrelevant, engineers can already read the signs and switch to backup equipment without any production loss.

Who is Juergen Weichenberger? An AI practitioner who has built industrial AI systems since the 1990s, led AI at Schneider Electric, and now serves as a data and AI partner at EY, focused on critical infrastructure including energy grids and heavy industry.

What happens when AI optimizes without constraints? It can find a mathematically optimal solution with almost no margin for error, or in commercial contexts, oversupply a market to the point of collapsing prices and profit margins, even while technically hitting the stated optimization target.

What are the four factors that actually drive value in industrial AI? Yield, energy, throughput, and quality. According to Weichenberger, these four fundamentals consistently outperform more heavily marketed AI use cases like predictive maintenance.

Watch the full conversation and explore more episodes at Unscripted with Jeff Pedowitz.