Three Things to Ask Before Leveraging AI in the Factory: What Siemens Says You Should Do Before Technology
Thank you for reading until the end. I will continue to organize trends in factory DX and smart factories from the perspectives of management, technology, and the shop floor. If you found this article helpful, I would appreciate your support through likes and follows.
There is almost no one left who isn't talking about AI.
However, when it comes to how to actually leverage it in a factory, the outline suddenly becomes vague.
After reading an article published in March 2026 by Chris Stevens, President of Siemens Digital Industries U.S., I wanted to organize three points of discussion to bridge that gap.
The title of the original article is "Three Key Industrial AI Insights for U.S. Manufacturing." Written based on discussions at the Federal Reserve Bank of Chicago's Automotive Industry Symposium, it focuses more on "what needs to be prepared beforehand" rather than flashy technical theories.
Stevens writes that when he asked the audience, "Who has talked about AI in the last two days?" almost everyone raised their hand. Interest is high. However, when it comes to how to realize it on the shop floor, the answer is not yet settled. I think this "coexistence of excitement and uncertainty" is a sentiment quite similar to that in the Japanese manufacturing industry.
First: Define "what you are trying to solve"
What Stevens puts first is not the talk of AI. What he wants to understand first is what you are trying to solve. Next is how the current process works. He says the order starts there.
This is understated, but I think it is a very important point. A factory is not a place where you line up the latest technology on a blank plot of land. Equipment is already running, people are supporting productivity through ingenuity, and you need to layer new mechanisms on top of that. Therefore, I think it is difficult for technology alone to provide value if you don't consider how things are currently running.
The article also states that conversations should always start here, and that it is important to identify where time is being lost, where quality is breaking down, and where flexibility is being lost. In other words, AI is a means to an end, not the goal itself. If you skip this order, it becomes easy for the question "What actually got better?" to become ambiguous after implementation. In that sense, it makes more sense to view digital twins not as a replacement for reality, but as a tool to understand the current process, verify the effects of changes on the virtual side, and lower the risk of trying them out on the shop floor.
Reading this reminded me of the "10-20-70 rule" I mentioned in this note before. It is the idea that the majority of results are determined by changes in people and processes, not by the technology itself.
Siemens' article also seems to be saying something quite similar in the end. Technology leads to value when the problems to be solved and the flow of the shop floor are articulated beforehand.
Second: AI only functions when it understands the "context of the entire factory"
The second point is data, but it is not just a simple matter of quantity. The Siemens article introduces a survey for manufacturers where 70% of responding companies thought they had "sufficient data," while "data quality" was cited as the biggest barrier to improvement.
This is quite suggestive. The data exists. But it is not connected in a usable form. Therefore, it does not lead to meaningful decisions within the factory. I think this is exactly the state that is likely to occur on the manufacturing floor.
Stevens expresses what people on the shop floor really want to do in a very easy-to-understand way. They want to stand in front of a machine and ask in natural language, "How was today's production?" and "Why did it drop by 10%?" However, for AI to answer that question properly, numerical values from a single piece of equipment are not enough. Only when the causal connections between motors, drive systems, robots, production lines, and the entire plant are visible does the data hold meaning.
There is a passage in the article to the effect that "dashboards tell you what happened, but context tells you why it happened." I think this is very important and quite succinctly expresses what is needed after visualization. Just lining up data is not enough; only when you can understand how that data connects to which process and where it affects the whole can AI output lead to meaningful decisions.
This feeling also overlaps with the talk of alert fatigue I wrote about before. Anomalies are being detected. Notifications are being sent. Yet the reason the shop floor doesn't move is not because there is no data, but because the context is missing. I think the problem is not quantity, but structure and connection.
Third: Future competitiveness will be "orchestration"
The third point, personally, gave me the most to think about. Mr. Stevens writes that many factories are built on a mix of technologies from different generations. There is old equipment beneath new software, vendors are fragmented, and years of operations have piled up on top of that. This must apply to quite a lot of sites.
Up to this point, it is a common story. But what was interesting was what came next. Analysis, machine learning, and AI are no longer confined to offline analysis. They are beginning to enter the realm of predicting maintenance timing during operation, optimizing production volume and throughput, and even providing real-time adjustment suggestions. In other words, it is based on the premise that the number of factories where individual intelligences operate simultaneously will increase.
When that happens, the problem is no longer "whether or not there is AI." The next point of discussion becomes whether the scheduling mechanisms, optimization engines, predictive maintenance models, and operator support applications will function without contradiction as a whole when they each operate intelligently. Even if individual systems are excellent, if there is no coordination, they will clash with each other, and in the end, the people on the front lines will bear the brunt of that.
Mr. Stevens expresses this state with the sentiment, "The problem is not that there is too much automation. It is uncoordinated automation." This is a very heavy statement. The value is not in increasing AI itself, but in being able to maintain overall safety, stability, and discipline even when multiple intelligences are operating simultaneously. This is where the orchestration mentioned in the article comes in—the idea of operating while coordinating multiple systems.
To be honest, I still do not have a clear picture of how to implement this orchestration on the factory floor. However, factories where multiple AI systems operate simultaneously will certainly increase from here on out. At that time, I feel that the problem of "being correct individually but not working well as a whole" will become a very realistic theme.
How to read this article
When you line up Mr. Stevens' three points, they are very straightforward. First, define what problem you are solving. Next, make visible how the factory's data and equipment are connected. Finally, operate while coordinating multiple intelligences.
Of course, this is also an article written from the perspective of a platform vendor, Siemens. There is no doubt that the worldview of being easy to connect to their own digital foundation, Xcelerator, and digital twins is in the background. Even so, I feel that the order of these three points itself has a very essential correctness.
In fact, at factory DX sites, it is not uncommon for discussions about tools to proceed while the "purpose of doing it" remains vague. You have visualized it, but the subsequent judgments and operations have not been decided. You have tried AI, but it has not been determined which tasks or where it will be utilized. When that happens, even if technology is introduced, it is difficult to lead to tangible results on the front lines.
In the end, the starting point for success is quite modest. Put the issues into words. Understand the current flow. Think in advance about what context exists there and which mechanisms will clash with each other. I think there are surprisingly many things to do before talking about AI.
Articles related to this theme
・Automation of major processes from 18% to 50%. Even so, what makes the difference was not technology.
・AI in manufacturing, more than half stop at the pilot stage. Why connected AI changes productivity
・Why do we overlook serious abnormalities even after promoting visualization?
Regarding consultations
I accept consultations as a personal activity regarding organizing points before introducing AI for manufacturing, organizing before consulting with AI vendors, and internal study sessions/brainstorming. For details, please see theself-introduction article.
If you use X, I also accept inquiries via DM.
Thank you for reading until the end.
In this note, I organize trends in factory AI, robots, and smart factories by connecting management, technology, and on-site implementation.
I track not only individual news introductions but also which companies enter factory DX from which entry point and where the difficulties of on-site implementation arise.
If you would like to continue reading about this theme in the future, I would be happy if you could follow me.
Sources/Reference Links
Siemens | "Three Key Industrial AI Insights for U.S. Manufacturing" |
https://www.siemens.com/en-us/company/insights/us-stories/industrial-ai-insights-us-manufacturers/