Automation is not a new topic since ChatGPT. Companies have been automating processes for decades, with macros, scripts, workflows, robotic process automation. It works, it's proven, and it has one decisive advantage: it's deterministic. If I calculate 3 plus 3 today, I get 6. Tomorrow too. The day after too.
The breakthrough of Large Language Models and advances in Natural Language Processing have massively changed expectations around automation. Suddenly everything seems possible. But this is exactly where the trap lies: not everything that is possible is also sensible.
Three Levels of Automation
In our consulting practice, we distinguish three levels. Each has its place, and its limits.
Level 1: Classic Automation, Without AI
Rule-based workflows, API integrations, database operations, file transfers, forms that are processed automatically. Everything that can be captured in clear if-then rules.
The big advantage: These processes are the most robust thing we can build. We operate entirely in the deterministic space. The result is predictable, reproducible, testable. No surprises.
Typical tools: n8n, Make, Zapier, Python scripts, REST APIs, classic RPA tools.
Level 2: AI Assistance, AI as a Tool in the Process
Here we add AI selectively to an otherwise classic process. Why? Because certain tasks cannot be solved with rules alone:
Natural language needs to be understood, classified, or generated. Images, handwriting, or scanned documents need to be recognized. Unstructured data needs to be converted into structured formats. Edge cases need to be identified that don't fit into simple rules.
At this level, the overall process remains pre-modeled: the flow is fixed. Only at specific points does AI come in as a tool to solve tasks that wouldn't be possible without it.
Level 3: AI Agents, AI as Decision-Maker
Here we flip the principle. Instead of pre-modeling the entire workflow, we provide an AI agent with tools, and the agent independently decides which ones to use, in what order, and how often.
The big advantage: Enormous flexibility. The agent can react to unforeseen situations, find creative solutions, and handle complex tasks for which we couldn't build a rigid workflow.
The downside: We give up control. And that requires guardrails, clear boundaries within which the agent may operate. Because an agent's autonomy also means: it can make wrong decisions. And the risks that arise from this must be actively limited.
Of course, there are many more automation levels and maturity models in the literature. But if you've understood these three levels, and that each has its very specific use case and value, then you've already come a long way.
A real-world example: we had a client running numerous automation processes, efficient, stable, reliable. Everything worked. Then came the request to add AI to these existing processes. Not because there was a concrete problem. Not because something was missing. But because AI happened to be the hot topic.
This is a pattern we see frequently. And it's exactly the wrong approach.
"We don't implement AI for the sake of AI. We implement AI where we need it to solve problems that can't be solved without it.", Thorsten Vellmerk
Why Less AI Is Often More
Artificial intelligence, as we know it today, is based on probabilities. This is fundamentally different from a calculator. With a calculator, 3 plus 3 is 6 today, 6 tomorrow, 6 always.
With probability-based systems, tomorrow the result might be 5.9. Or 6.1. These deviations are inherent to the system, and in many cases even desirable, because they enable the creativity and flexibility that makes AI so powerful.
Of course, we work to minimize these deviations during development, through better prompts, tighter contexts, validation steps. But the principle remains:
"If I can solve something deterministically, I don't use AI. Because every AI component makes the process a bit less certain, and usually more expensive.", Thorsten Vellmerk
This is not an anti-AI statement. On the contrary: it's the prerequisite for deploying AI where it truly shines. If we use AI everywhere, even where a simple if-then suffices, we dilute its value while simultaneously increasing cost, complexity, and error susceptibility.
Choosing the Right Level
In our consulting, we always start with the question: At which level should this process be automated? The answer is often surprising: many processes that clients believe require AI can be solved with classic automation, faster, cheaper, more reliably.
And where AI is truly needed, we deploy it deliberately, as assistance in a controlled process or as an agent with clearly defined guardrails.
Conclusion
The three levels of automation are not a progression model where you want to reach the highest level as fast as possible. They are a toolbox. And as with any good toolbox: use the tool that fits the problem, not the one that looks most impressive.
Not sure which level is right for your process? Get in touch, we'll help you find the right balance between automation and AI.