AI has arrived in SMEs, but mostly for language
The use of artificial intelligence in Swiss SMEs has risen sharply in a short time. According to the 2025 SME labour market study, the share of companies that deliberately integrate AI into their work processes grew from 22 to 34 percent within a single year. The most common uses are language tasks: translations at 52 percent, correspondence at 47 percent.
This reveals something important: today, AI mainly helps with wording. The actual logic of a business process, meaning what happens when and with which data, is a different question. And it is precisely there, inside an automation, that a decision arises which determines effort and reliability: does this step really need AI, or is a fixed rule enough?
Two tools, two strengths
A fixed rule is a clear instruction following the pattern „if this, then that“. Given the same input, it always produces the same result. It is cheap to run, predictable, and traceable when something goes wrong, because every step can be followed back. An example: when an invoice arrives in a fixed format, the system reads the same fields every time and forwards it to the right place. This runs identically a thousand times over.
An AI component works differently. It can handle text that has no fixed format, it classifies and makes judgements. It recognises that two differently worded emails carry the same request, or pulls the right detail from a PDF even when its layout changes each time. From our experience, it can be summed up like this: AI components in automation are ideally suited where decisions and interpretation are involved. Where a person would otherwise have to take a quick look and classify something, AI is strong.
How to tell what your step needs
The question is not „AI or not“ for the whole process, but for each individual step within it. A fixed rule is usually enough when the input has a clear format and the process follows clear rules:
- Transferring the fields of a delivery note into the inventory system, when the layout is always the same.
- Capturing and booking a shift report from production.
- Checking and forwarding an invoice with a known schema.
An AI component is worth it when the input is unstructured or calls for a judgement:
- Assigning emails without a fixed format to the right department.
- Categorising a free-text complaint.
- Extracting the required detail from documents with varying layouts.
As a rule of thumb: where structure and fixed sequences are involved, the rule is the better choice. Where understanding and classifying are involved, AI comes into play.
The price of flexibility: AI needs more upkeep
The flexibility of AI comes at a price, and it would be dishonest to hide it. A fixed rule runs unchanged until the process itself changes. An AI component, by contrast, can change without anyone touching it.
There are two reasons for this. First, providers update their models continuously, sometimes without notice. The same call that worked correctly yesterday may return a slightly different result tomorrow. Second, AI is sensitive to the exact wording of the instruction: even a small change in wording can shift its behaviour disproportionately. Neither of these exists with a fixed rule. Code does not change by itself between Tuesday and Wednesday.
In practice this means an AI component needs ongoing oversight. Its output has to be monitored and checked so that a silent change does not slip through unnoticed. A fixed rule barely needs this. That is not an argument against AI. But it is an effort that has to be planned for from the start, and that belongs on the scale when asking „is AI worth it for this step“.
The best approach is often the combination
In most cases it is not an either-or. The most solid solution combines both: fixed rules handle the structure and the checking, the AI component handles only the one step in between where understanding is required. And the output of the AI is then checked by fixed code before it moves on. This way you use the strength of AI without letting its unpredictability run unchecked through the process.
This matches our approach of choosing individual steps deliberately. As we described in our article on local automation, we decide for each step individually what happens where. The question of AI is part of the same logic: not deploying the most modern option, but the fitting one.
How we decide
When we build an automation, we ask at every step first whether a fixed rule can solve it. If it can, we use it, because it is cheaper and more reliable. Only when a step requires genuine understanding or a judgement do we add an AI component, and then with the necessary oversight around it.
Are you unsure whether your project needs AI or whether a fixed rule is enough? We will look at your process together and tell you honestly what makes sense and what does not. The first conversation is free of charge.
