Scaling Agentic AI Workflows in Manufacturing with Causal AI

Most manufacturers already have the data. They have historians, SCADA systems, data lakes, and maybe even a unified namespace. But when they try to deploy agentic AI on top of all that infrastructure, the agents can't actually reason about what's happening on the shop floor—because the data has no meaning.

‍

In my conversation with Bernhard Kratzwald, co-founder and CTO of Ethon AI, on the AI in Manufacturing podcast, he made a case that stopped me cold: the bottleneck for agentic AI in manufacturing isn't data access or compute power. It's context. Specifically, it's the absence of a living, causal knowledge graph that mirrors how production actually works. Without that layer, even the most sophisticated AI models are flying blind—and manufacturers keep cycling through pilots that never scale.

‍

‍

Why Is Manufacturing's Existing Data Stack Failing Agentic AI?

‍

The industrial data stack was never designed for the questions that agentic workflows need to ask. Legacy historians store time series data. Data lakes accumulate batch exports from SAP and quality systems. A unified namespace gives you hierarchical tag structures so you know which sensor belongs to which line. All of that is necessary. None of it is sufficient.

‍

As Bernard put it, with a UNS you can ask "give me all the tags that measure temperature." That's useful. But an agentic workflow needs to ask something fundamentally different: "Why did this batch require extensive rework today?" Answering that question requires knowing which tank the batch was in, what time it moved between tanks, what recipe it was manufactured with, what interventions operators made, and what the time series looked like at each stage. That kind of reasoning demands semantic relationships between data points—not just a well-organized tag structure.

‍

The gap between what the data stack provides and what agentic AI requires is where most implementations stall. You pull all the data into a root cause investigation and discover the data is either disconnected, lacks context, or is simply wrong because nobody was watching it closely enough to catch calibration drift. This is the "context gap"—and it's the single biggest reason manufacturing lags behind other industries in deploying agentic workflows at scale.

‍

What Changed That Makes Traditional Continuous Improvement Insufficient?

‍

The objectives haven't changed. Manufacturers still need to produce more, cut costs, and defend margins. The old wisdom—that twenty cents of every manufacturing dollar goes to waste—was true fifty years ago and will probably be true fifty years from now, because the frontier keeps moving.

‍

What has changed is that traditional methods have maxed out. Automation has been pushed about as far as it can go. Classic data science and machine learning have been applied across the board for decades. If you want to capture the next twenty percent of improvement, you need methods that can reason about why things happen, not just detect that something happened.

‍

Meanwhile, complexity has exploded. Manufacturers that ran ten or fifteen SKUs a decade ago now run fifty or a hundred. Cost pressure has intensified. The workforce is turning over, and the process engineers who spent thirty years learning every quirk of a production line are retiring. Bernard described a reality where some manufacturers need five years to train a process engineer to full fluency. That timeline is incompatible with current market demands. The knowledge walking out the door can't be replaced fast enough by traditional means.

‍

What Is the Real Cost of Running AI Without Causal Context?

‍

The failure pattern Bernard described is one I see constantly: manufacturers invest heavily in data infrastructure, deploy an AI model, and then watch it produce recommendations that operators ignore. Not because the operators are resistant to change, but because the AI can't explain why it's making a suggestion—and the suggestion often doesn't account for the physical, mechanical, and procedural realities of the shop floor.

‍

This is what I'd call the "correlation trap." Correlation-based analytics can tell you that production runs better in winter than summer. That's interesting but completely unactionable—you can't shut down the factory for seven months. Causal AI, by contrast, can trace the chain: in winter, feeding water temperature is lower, which changes cooling behavior, which improves a specific quality attribute, which means less rework. Now you have something you can act on. Maybe in summer you need to pre-cool the water to replicate winter conditions.

‍

Without that causal chain, every AI recommendation is essentially a black box. And in a safety-critical environment—where incorrect temperature setpoints can endanger equipment, workers, or consumers of pharmaceutical products—a black box isn't just unhelpful. It's a liability. One previous guest on this podcast made the point that "explainable AI" as typically implemented wouldn't hold up in court, but a documented chain of causal reasoning would. That distinction matters enormously in regulated industries.

‍

The compounding cost is subtle but devastating: every failed pilot erodes organizational trust in AI. Every recommendation an operator ignores reinforces the belief that these tools don't understand the real world. And every month spent in pilot purgatory is a month where the knowledge gap widens as experienced workers leave.

‍

How Does a Process Knowledge Graph Change Agentic AI in Manufacturing?

‍

The shift Bernard advocates isn't about adding another tool to the stack. It's a fundamentally different approach to how data relates to itself.

‍

A process knowledge graph is, at its technical core, an ontology—a set of triples where two objects are connected by a concept. This tank contains these seventeen tags. This tag measures quality. This tank is connected to that tank, meaning liquid can flow between them. This batch was manufactured with this recipe. Each triple is simple. The emergent capability is not.

‍

Once you have those semantic relationships in place, you can ask questions that were previously impossible without a human expert manually pulling data from six different systems. You can ask which five machines in a factory caused the most downtime today. You can connect ten factories to the same conceptual model and ask why one line outperforms another. You can run a root cause investigation that automatically pulls every relevant data point from every system—time series, batch records, recipes, operator interventions, maintenance logs—and feeds it to a causal model that identifies not just what correlated with the problem, but what caused it.

‍

The critical insight from Bernard's approach is that this knowledge graph doesn't need to be built all at once. In fact, he's skeptical of the "build the operating system first" approach, where you try to model everything upfront before deploying any applications. Instead, Ethon builds the knowledge graph incrementally, driven by specific use cases that deliver ROI. You start with a root cause investigation workflow. The knowledge graph gets built to the depth needed to support that workflow. Insights from the investigation get fed back into the graph. The graph grows richer over time, and each new use case extends it further.

‍

This is context as a living thing—not a one-time modeling exercise. As Bernard said, you're never done building a knowledge graph, because there's always more knowledge you can distill. The question is when the marginal returns stop justifying the effort for a given use case.

‍

How Fast Can Manufacturers Deploy Agentic AI and See Results?

‍

Ethon restructured their entire deployment motion over the past eighteen months, adopting a forward-deployed engineering model. Their teams spend two to three days up to a week and a half on-site with customers, understanding processes, helping with data pipelining, building the initial process model, and training users.

‍

The result: time to first value is now consistently below three months from project kickoff. That timeline includes not just technical deployment but the change management work that Bernard considers non-negotiable. Understanding how an operator physically executes a step, recognizing that adding even one second to a repeated task compounds across thousands of repetitions, designing interfaces that work for someone who's been on the floor for five hours or fifty years—all of that is part of the deployment, not an afterthought.

‍

The architecture itself is modular. If a manufacturer already has a UNS, Ethon sits on top of it. If they already have a data lake, same. If they have nothing, Ethon can assist or recommend partners for the infrastructure layer. The platform ingests data via MQTT, Kafka, Sparkplug for streaming, or batch-based APIs for SAP exports and SQL databases. It runs as SaaS in the cloud, with private cloud options for large enterprises and edge-compatible modules for latency-sensitive applications like optical inspection.

‍

The proof is in the scale: Siemens, one of Ethon's earliest customers, has deployed across twenty-plus factories with documented savings of approximately ten million US dollars, recently featured in a World Economic Forum publication. Lindt & Sprüngli has achieved considerable waste reductions globally. These aren't pilot results. They're production-scale outcomes.

‍

What Question Should Manufacturing Leaders Actually Be Asking About AI?

‍

The wrong question is "how do we implement agentic AI?" The right question is "how do we make every operator as effective as our most experienced one—and keep getting better after that person retires?"

‍

That reframe changes everything about how you evaluate solutions. It shifts the focus from technology selection to knowledge capture and operationalization. It makes the knowledge graph not an infrastructure project but a strategic asset. And it makes the timeline urgent, because the expertise walking out the door doesn't wait for your data platform to be perfect.

‍

Bernard's prediction for the next five to six years centers on three shifts: assisted workflows becoming autonomous across entire value streams, cross-factory intelligence becoming standard rather than aspirational, and workforce transformation accelerating to the point where new operators achieve fluency in months instead of years. The manufacturers who start building their causal knowledge infrastructure now will be the ones positioned to capture those gains. The ones who wait will find the gap has only widened.

‍

As Bernhard put it: the only mistake you can make today is not doing anything. The best time to start was yesterday. The second best time is today.

‍