AI & Automation

Foundations and context for industrial practice

We use AI to create technical specifications, support programming logic (as far as Siemens software allows), produce manuals and operating guides, and build test environments for our programs. Integration into live plants, especially for anomaly detection, is in preparation.

Quick access

How does an AI actually work?

Artificial intelligence follows one basic principle: a model is trained with very large volumes of data and learns to recognize patterns and predict probabilities. Different inputs create different models:

• Language models (LLMs): process text, understand relations between words and sentences, and generate answers, translations or code.
• Image models: some generate images from descriptions, others detect objects and structures in existing images.
• Audio models: convert speech to text, generate speech from text, or create music and sounds.
• Video models: create clips or detect and analyze motion and objects in video streams.

All models share the same workflow: an input arrives (text, image, audio). An invisible instruction, often called the system prompt, sets the role or context. The neural network processes the data, detects patterns and calculates probabilities. The output is a human-readable result – text, image, speech, music or analysis.

The takeaway: AI does not think; it calculates the most likely fit. The same principle powers systems that write text, design images, understand speech, generate audio or analyze videos.

What is a system prompt and why is it important?

Beyond training data, the system prompt is central. It is the invisible start command that tells the AI which role or behavior to adopt.

The same language model can produce very different results depending on the prompt. One sentence changes dramatically by role:

• Scientist: “Sunset is the result of Earth’s rotation, where the observer perceives the sun dropping below the local horizon.”
• Poet: “The sun sinks in purple glow and kisses the horizon goodbye.”
• Teacher: “The sun shines during the day. In the evening it disappears behind the horizon, then it gets dark.”
• Programmer:

  def sunset():
      return "The sun sets in the evening."
  

• Manager: “The daily sunset marks a natural deadline after which resource planning should shift to the night phase.”

So the prompt steers the output even though the underlying model is the same. In practice, the right prompt lets an AI answer factually, creatively, legally, technically or entertainingly – whichever perspective is needed.

Example system prompt

You are an experienced automation engineer.
Your job is to write technical documentation clearly,
with a structured outline and in line with standards.
Use precise terminology, a consistent structure
and a factual tone.

What is a context window and what are hallucinations?

For useful answers, an AI must “see” what is being discussed. The current conversation or text does not fit into the model without limits; instead there is a context window – the working area that holds the latest sentences, paragraphs or pages.

Everything inside this window can be used directly. Information that falls outside or has already been pushed out is no longer visible. The model does not remember like a human; it only processes what is currently in the window.

If a question arrives without matching context, the AI may still answer by falling back on patterns from training data. This is where hallucinations arise: outputs that sound plausible but are factually wrong or invented.

• Context window: the span the AI can observe at once.
• Hallucination: when the model lacks solid clues yet still produces something that looks like a correct answer.

For users this means: the better you control context – e.g., via precise input or additional documents – the more reliable the answer becomes.

Can we extend an AI’s “awareness”?

Strictly speaking an AI has no awareness. It computes probabilities and recognizes patterns but knows nothing in the human sense. Still, there are ways to extend its usable knowledge and capabilities:

• RAG (Retrieval-Augmented Generation): the model is supplied with additional material during the query – documents, databases, domain texts. It combines language capability with external information to access current or highly specific content not in the original training data.
• LoRA (Low-Rank Adaptation): small adapter modules are added to a large model to adapt it to new styles or tasks. For language models this enables lightweight fine-tuning to specialize on a domain such as medicine, law or engineering.
• Fine-tuning / continued training: an LLM can be trained further with new examples to answer more precisely in a field. This is more effort than RAG or LoRA but allows deep specialization.

In short: you cannot expand an AI’s consciousness, but you can extend its knowledge context and skills via external retrieval, light adapters or additional training.

Can an AI model replace a PLC?

Short answer: no, not in the foreseeable future.

A PLC performs tasks that differ fundamentally from what an AI model can do. A controller is deterministic, runs its cycles precisely in milliseconds and always delivers the same result. That reliability is crucial when industrial processes must run predictably.

Controllers are also bound to strict norms and safety requirements. Industrial systems must be provably safe, validatable and certifiable. An AI model is probabilistic: it calculates likelihoods and can vary its output with context. That is valuable for analysis but does not meet the demands of a safety-critical real-time system.

Robustness differs as well. Controllers are built to run for decades under harsh conditions. Large AI models run on powerful hardware and need regular updates and maintenance.

That does not mean AI has no place in automation. It can complement controllers for anomaly detection or predictive maintenance to spot deviations early. It can optimize energy use or dynamically adjust production plans. It can also support engineers and technicians by drafting function blocks or triaging error messages quickly.

Bottom line: a PLC is a safety-critical real-time system. An AI model is a powerful pattern recognizer and optimizer. They are not interchangeable but can work effectively together.

Is working with AI safe and what about data protection?

The answer depends on how AI is deployed. Three common setups:

• Browser chatbots: services like ChatGPT, Claude, Gemini or Grok run entirely in the browser. All data travels over the internet and lives outside your infrastructure. Convenient, but all traffic is processed by external providers.
• API integration: the model runs in the cloud, while key controls such as system prompts, RAG sources or internal databases stay in your network. That gives more control over what data leaves the company.
• Local operation: models are installed on your own machines and run fully offline. This offers maximum data security, but requires significant compute and know-how to deploy and maintain.

So safety is driven less by the technology itself and more by the operating model. From cloud chatbots through hybrid API setups to fully offline operation, there are tiers to match your privacy and infrastructure requirements.

What are agents and what do they do?

An agent extends a language model so it can act, not just answer. While an LLM mainly understands and generates text, an agent adds action logic and tools.

• Language model core: understands inputs and prepares decisions.
• Context or memory: keeps the agent consistent across multiple steps.
• Tools and interfaces: enable external actions such as database queries, starting programs or reading sensor values.
• Control logic: decides which action the agent picks and when.

This enables automation beyond text generation. An agent can gather information from different sources, plan intermediate steps, prepare results and then execute an action. In automation practice, agents can monitor systems, analyze data or assist engineers with automatic suggestions and documentation.

So an agent is not a new model, but the surrounding structure that embeds a model into concrete workflows and takes over practical, repeatable tasks.

Continue the conversation

Want to explore concrete use cases or evaluate a solution? Talk to us.