We've stopped asking each other
The lecture began with a graph that caught many people off guard. On Stack Overflow, the forum where programmers have been helping each other for nearly 20 years, the number of new questions has dropped by 99 percent since ChatGPT was launched. People haven’t stopped having problems. They’re asking AI instead.
The same thing happens at the factory. The new engineer might ask the chatbot instead of an experienced colleague. That raises the question of whether the chatbot understands your specific process.
AI is not a thing
There is no single definition of AI. The term encompasses everything from rules written by humans to systems that teach themselves. The lecture clarified the difference between machine learning, deep learning, and large language models (LLMs).
The difference is practical. A typical program follows rules that someone has defined in advance—for example, triggering an alarm if the temperature exceeds 80 degrees. A model based on machine learning, on the other hand, teaches itself—for example, from images of intact and broken products. A language model does something else entirely: it generates text by predicting the next word.
Asking for production data and predicting a machine failure are, therefore, two completely different tasks that require different solutions. The fact that a product “has AI” therefore says very little about what it actually does.
I've read a lot about it, but I've never been to your factory
A language model like Meta’s Llama 3 has been trained on fifteen trillion word units. That’s equivalent to more than 3,000 human lifetimes of reading. Yet it knows nothing about your specific process unless someone tells it.
And when it doesn’t know, it guesses anyway, with the same confident tone. An example from the lecture: ask for the cleaning instructions for a specific tank, and you might be given a document number and a version that sound perfectly reasonable—but don’t actually exist. That’s called a hallucination.
In production, and especially in GMP environments, it must therefore always be possible to see where a result comes from. Instruction files can control how the model works and where it searches, but they never replace verification of the result.
From Chatbot to Agent
A chatbot responds; an agent takes action. An agent works in a loop: it plans, uses tools, reviews the results, and goes through the process again if the goal hasn’t been achieved. The lecture itself was an example of this, since it was created with the help of an AI agent that searched for sources, generated images, and corrected its own errors.
For an agent to be useful in the factory, it must be able to access your systems. That’s where the Model Context Protocol (MCP) comes in. The presentation compared it to USB-C: a common interface between AI and, for example, MES or ERP systems, which can often be built on top of existing APIs. There are already over 10,000 MCP servers.
The technical issue will soon be resolved. The important question is what the agent will be allowed to do. Reading data is one thing. Changing a recipe or a setpoint is something entirely different.
That's why pilots get stuck
Price is rarely the obstacle. The same level of performance costs about 280 times less today than it did two years ago. What brings AI projects down is the foundation.
Ver Takeo shared his own experience from a predictive maintenance project. The architecture was built while the pilot was underway; the process was unstable, and the data needed to train the model could not be obtained. Without a stable baseline, there was nothing for the model to learn.
This tied in with today’s keynote on digital maturity: you can’t skip steps. Prediction requires that you first have visibility and understand your process. But not all applications need to wait that long. An AI assistant that helps operators search through documentation or translates error codes into understandable Swedish can be useful much sooner. Even in that case, the answers must be verifiable.
Cloud or on-premises, and who is responsible for this issue?
The cloud offers the best models and a quick start, but the data leaves the company. On-premises servers provide control and work without an internet connection, but require in-house management and offer less powerful models. There is no right answer, but the question needs to be asked early on.
In the discussion that followed, Cecilia Jacobsson highlighted the issue of ownership. Local pilot projects with weak architecture and unclear lines of responsibility make it difficult to maintain a holistic view. Her message was to manage AI with clear lines of responsibility and clear processes, just as is done in the rest of the IT environment.
Four Questions to Start With
- Is the right data available, and can it be accessed?
- Is the process stable enough to learn from?
- How do we verify that the answers are correct?
- Who owns the solution once the pilot is complete?
The answers will determine whether a limited information support system is the right first step, or whether data and architecture need to come first.
Would you like to discuss further how AI can benefit your production? Get in touch, and together we'll find the next step that fits where you are today.