It is common sense that preventing problems from recurring is better than sweeping them under the rug, yet most organizations seem to prefer the rug. Some industry leaders like Lindt & Sprüngli and Bosch, however, use new AI capabilities to track down and remove hidden inefficiencies.
The limits of classic RCA tools
Classic root-cause analysis tools can be quick and useful. For example, filling a cause-and-effect fishbone diagram or using the “5-whys” method of asking “Why did that happen?” several times until the answer reveals the root cause. A group of experienced people can use these tools to brainstorm solutions and implement effective countermeasures targeting the root cause. Yet, these methods quickly meet their limits. If the cause is unobservable or multifaceted, the correct answer to the next “why” question should be, “I don’t know.”
For more advanced analysis, statistical process control (SPC) has been the go-to method for a century. SPC monitors a process by tracking assumed mission-critical process parameters and output metrics. If all variables are observed to be within the upper and lower limits, the process is considered “in control” and is expected to produce consistent, high-quality parts. The fundamental idea is to identify and reduce process variability to maintain stability and predictability.
But SPC also has its limitations: establishing and maintaining SPC is time- and resource-intensive, and even when all variables are within specification limits, there is no guarantee that the output is [1]. This method, more than a hundred years old, falls short in complex, integrated production lines, where nonlinear relationships among variables can significantly affect output. Today, there is a better way.
Learning to see hidden relationships
Artificial intelligence has created new possibilities for root cause analysis that were previously out of reach. A common myth in manufacturing is that AI cannot pinpoint root causes because it operates in a “black box” fashion and only understands correlations. However, recent developments in AI have significantly improved interpretability and explainability, allowing engineers to better understand how different process variables relate to outcomes [2].
When coupled with process models, human domain expertise, and physical constraints in knowledge graphs, the AI models become very powerful investigators. Rather than relying solely on correlations, AI can now uncover previously unknown causal relationships [3].
Consider, for example, Lindt & Sprüngli’s chocolate production:
Lindt & Sprüngli operates 12 production facilities across Europe and the United States. Chocolate manufacturing process is riddled with nonlinear and convoluted relationships. For example, the temperature during an early mixing process may affect the viscosity of the chocolate mass in a transfer pipe, making it more sticky during the molding process, but only when the molding pressure falls within a narrow range. How can Lindt & Sprüngli reduce such complex yield loss? Lindt & Sprüngli uses Ethon’s industrial AI platform that allows it to identify and address hidden root causes. By connecting and contextualizing input and output data from machinery, sensors, and software systems, Lindt & Sprüngli benefits from knowledge graphs to uncover previously unknown relationships and boost yield performance.
The root cause analysis illustrated in the chocolate example occurs within a production line, but AI does not have to stop at the factory gate. A case in point can be learned from automotive supplier giant Robert Bosch:
If doing root cause analysis is challenging within a factory, imagine doing it across a manufacturing network. Robert Bosch GmbH can identify root causes of productivity problems that originate in upstream suppliers. In an awarded example, Bosch used Ethon’s industrial AI platform to analyze process data across a supply chain, identifying machining settings associated with downstream quality issues. The relationship was previously unknown to Bosch and would have gone unnoticed without the AI system. Identifying and eliminating such complex root causes enhances quality, minimizes waste and CO2 emissions, and reduces costs in Bosch’s production of automotive parts.
But if AI is so great at root cause analysis, what’s keeping other companies away from using it?
The 3C data strategy
Given the potential benefits, companies are scrambling to adopt AI. But they often realize that they are not AI-ready. For years, companies did not have to pay much attention to maintaining data beyond the bare minimum. Other collected data was erased or dumped in its native format onto a storage medium or platform. They had no factory data strategy.
A foundational step for AI-powered root cause analysis is to follow the 3C strategy: collect, connect, and contextualize representative data across the factory and beyond. First then can the process variables be related to the output quality and reveal complex and non-trivial relationships as in the above examples of Lindt & Sprüngli and Bosch.
Many managers choose a more convenient but ineffective approach to AI scaling. Because some machines generate gigabytes or terabytes of data, the managers often claim to be “data-rich” and seek to make use of it. Drawing on the most available data pipelines, typical AI applications calculate the remaining useful life of modern machinery or identify more efficient ways to manage energy consumption in facilities.
The problem with a push approach to AI applications is that it doesn’t start from a mission-critical problem. For example, it is fine if a modern machine can produce more parts per hour through AI-based asset monitoring, but it may have a limited effect on the bottom line if it isn’t a bottleneck in the production line.
From data to aha!
To make good use of AI, companies must prioritize addressing their core business challenges. And, in manufacturing, the real problems are more likely to stem from horizontal flows across multiple processes than from a single data-spouting machine. From there, define what data can help model the process flows, invest in the needed infrastructure, and start collecting representative data. Once the data is available, it must be made interpretable within the same context by connecting it to relevant metadata, such as locations, product or batch identities, or timestamps.
Since AI-based root cause analysis requires an ensemble of data across a vast ecosystem of OT/IT infrastructure, it is not a bolt-on app to existing software. It benefits from tailor-made software that extracts relevant data from disparate sources, contextualizes it within a representative process model, and uses tailored AI models to analyze it [4].
The richer and more representative the data, the better the decision support will be for the users. But only when companies collect, connect, and contextualize that data can they identify hidden root causes of problems with the help of AI.
References
- Baudin, M., & Netland, T. (2022). Chapter 16 Improving Performance, Introduction to Manufacturing: An Industrial Engineering and Management Perspective. Taylor & Francis.
- Senoner, J., T. Netland, and S. Feuerriegel, Using Explainable Artificial Intelligence to Improve Process Quality: Evidence from Semiconductor Manufacturing. Management Science, 2022. 68(8): p. 5704-5723.
- Feuerriegel, S., Y.R. Shrestha, and G. Von Krogh, A New Machine Learning Approach Answers What-If Questions. MIT Sloan Management Review, 2025. 66(3): p. 65-69.
- Senoner, J., Kratzwald, B., Philippsen, R., & Netland, T. (2024). Manufacturing analytics system: A new IT category enabling next-level operational excellence. IFAC-PapersOnLine, 58(19), 831-835.
What's on your mind?